xref: /illumos-gate/usr/src/uts/common/os/strsubr.c (revision dc7ec32189c86a0f330aee77229dad2ad57eac71)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T	*/
22 /*	  All Rights Reserved	*/
23 
24 
25 /*
26  * Copyright 2010 Sun Microsystems, Inc.  All rights reserved.
27  * Use is subject to license terms.
28  * Copyright (c) 2016 by Delphix. All rights reserved.
29  * Copyright 2018 OmniOS Community Edition (OmniOSce) Association.
30  * Copyright 2018 Joyent, Inc.
31  * Copyright 2022 Garrett D'Amore
32  * Copyright 2025 Oxide Computer Company
33  */
34 
35 #include <sys/types.h>
36 #include <sys/sysmacros.h>
37 #include <sys/param.h>
38 #include <sys/errno.h>
39 #include <sys/signal.h>
40 #include <sys/proc.h>
41 #include <sys/conf.h>
42 #include <sys/cred.h>
43 #include <sys/user.h>
44 #include <sys/vnode.h>
45 #include <sys/file.h>
46 #include <sys/session.h>
47 #include <sys/stream.h>
48 #include <sys/strsubr.h>
49 #include <sys/stropts.h>
50 #include <sys/poll.h>
51 #include <sys/systm.h>
52 #include <sys/cpuvar.h>
53 #include <sys/uio.h>
54 #include <sys/cmn_err.h>
55 #include <sys/priocntl.h>
56 #include <sys/procset.h>
57 #include <sys/vmem.h>
58 #include <sys/bitmap.h>
59 #include <sys/kmem.h>
60 #include <sys/siginfo.h>
61 #include <sys/vtrace.h>
62 #include <sys/callb.h>
63 #include <sys/debug.h>
64 #include <sys/modctl.h>
65 #include <sys/vmsystm.h>
66 #include <vm/page.h>
67 #include <sys/atomic.h>
68 #include <sys/suntpi.h>
69 #include <sys/strlog.h>
70 #include <sys/promif.h>
71 #include <sys/project.h>
72 #include <sys/vm.h>
73 #include <sys/taskq.h>
74 #include <sys/sunddi.h>
75 #include <sys/sunldi_impl.h>
76 #include <sys/strsun.h>
77 #include <sys/isa_defs.h>
78 #include <sys/pattr.h>
79 #include <sys/strft.h>
80 #include <sys/fs/snode.h>
81 #include <sys/zone.h>
82 #include <sys/open.h>
83 #include <sys/sunldi.h>
84 #include <sys/sad.h>
85 #include <sys/netstack.h>
86 
87 #define	O_SAMESTR(q)	(((q)->q_next) && \
88 	(((q)->q_flag & QREADR) == ((q)->q_next->q_flag & QREADR)))
89 
90 /*
91  * WARNING:
92  * The variables and routines in this file are private, belonging
93  * to the STREAMS subsystem. These should not be used by modules
94  * or drivers. Compatibility will not be guaranteed.
95  */
96 
97 /*
98  * Id value used to distinguish between different multiplexor links.
99  */
100 static int32_t lnk_id = 0;
101 
102 #define	STREAMS_LOPRI MINCLSYSPRI
103 static pri_t streams_lopri = STREAMS_LOPRI;
104 
105 #define	STRSTAT(x)	(str_statistics.x.value.ui64++)
106 typedef struct str_stat {
107 	kstat_named_t	sqenables;
108 	kstat_named_t	stenables;
109 	kstat_named_t	syncqservice;
110 	kstat_named_t	freebs;
111 	kstat_named_t	qwr_outer;
112 	kstat_named_t	rservice;
113 	kstat_named_t	strwaits;
114 	kstat_named_t	taskqfails;
115 	kstat_named_t	bufcalls;
116 	kstat_named_t	qremoved;
117 	kstat_named_t	sqremoved;
118 	kstat_named_t	bcwaits;
119 	kstat_named_t	sqtoomany;
120 } str_stat_t;
121 
122 static str_stat_t str_statistics = {
123 	{ "sqenables",		KSTAT_DATA_UINT64 },
124 	{ "stenables",		KSTAT_DATA_UINT64 },
125 	{ "syncqservice",	KSTAT_DATA_UINT64 },
126 	{ "freebs",		KSTAT_DATA_UINT64 },
127 	{ "qwr_outer",		KSTAT_DATA_UINT64 },
128 	{ "rservice",		KSTAT_DATA_UINT64 },
129 	{ "strwaits",		KSTAT_DATA_UINT64 },
130 	{ "taskqfails",		KSTAT_DATA_UINT64 },
131 	{ "bufcalls",		KSTAT_DATA_UINT64 },
132 	{ "qremoved",		KSTAT_DATA_UINT64 },
133 	{ "sqremoved",		KSTAT_DATA_UINT64 },
134 	{ "bcwaits",		KSTAT_DATA_UINT64 },
135 	{ "sqtoomany",		KSTAT_DATA_UINT64 },
136 };
137 
138 static kstat_t *str_kstat;
139 
140 /*
141  * qrunflag was used previously to control background scheduling of queues. It
142  * is not used anymore, but kept here in case some module still wants to access
143  * it via qready() and setqsched macros.
144  */
145 char qrunflag;			/*  Unused */
146 
147 /*
148  * Most of the streams scheduling is done via task queues. Task queues may fail
149  * for non-sleep dispatches, so there are two backup threads servicing failed
150  * requests for queues and syncqs. Both of these threads also service failed
151  * dispatches freebs requests. Queues are put in the list specified by `qhead'
152  * and `qtail' pointers, syncqs use `sqhead' and `sqtail' pointers and freebs
153  * requests are put into `freebs_list' which has no tail pointer. All three
154  * lists are protected by a single `service_queue' lock and use
155  * `services_to_run' condition variable for signaling background threads. Use of
156  * a single lock should not be a problem because it is only used under heavy
157  * loads when task queues start to fail and at that time it may be a good idea
158  * to throttle scheduling requests.
159  *
160  * NOTE: queues and syncqs should be scheduled by two separate threads because
161  * queue servicing may be blocked waiting for a syncq which may be also
162  * scheduled for background execution. This may create a deadlock when only one
163  * thread is used for both.
164  */
165 
166 static taskq_t *streams_taskq;		/* Used for most STREAMS scheduling */
167 
168 static kmutex_t service_queue;		/* protects all of servicing vars */
169 static kcondvar_t services_to_run;	/* wake up background service thread */
170 static kcondvar_t syncqs_to_run;	/* wake up background service thread */
171 
172 /*
173  * List of queues scheduled for background processing due to lack of resources
174  * in the task queues. Protected by service_queue lock;
175  */
176 static struct queue *qhead;
177 static struct queue *qtail;
178 
179 /*
180  * Same list for syncqs
181  */
182 static syncq_t *sqhead;
183 static syncq_t *sqtail;
184 
185 static mblk_t *freebs_list;	/* list of buffers to free */
186 
187 /*
188  * Backup threads for servicing queues and syncqs
189  */
190 kthread_t *streams_qbkgrnd_thread;
191 kthread_t *streams_sqbkgrnd_thread;
192 
193 /*
194  * Bufcalls related variables.
195  */
196 struct bclist	strbcalls;	/* list of waiting bufcalls */
197 kmutex_t	strbcall_lock;	/* protects bufcall list (strbcalls) */
198 kcondvar_t	strbcall_cv;	/* Signaling when a bufcall is added */
199 kmutex_t	bcall_monitor;	/* sleep/wakeup style monitor */
200 kcondvar_t	bcall_cv;	/* wait 'till executing bufcall completes */
201 kthread_t	*bc_bkgrnd_thread; /* Thread to service bufcall requests */
202 
203 kmutex_t	strresources;	/* protects global resources */
204 kmutex_t	muxifier;	/* single-threads multiplexor creation */
205 
206 static void	*str_stack_init(netstackid_t stackid, netstack_t *ns);
207 static void	str_stack_shutdown(netstackid_t stackid, void *arg);
208 static void	str_stack_fini(netstackid_t stackid, void *arg);
209 
210 /*
211  * run_queues is no longer used, but is kept in case some 3rd party
212  * module/driver decides to use it.
213  */
214 int run_queues = 0;
215 
216 /*
217  * sq_max_size is the depth of the syncq (in number of messages) before
218  * qfill_syncq() starts QFULL'ing destination queues. As its primary
219  * consumer - IP is no longer D_MTPERMOD, but there may be other
220  * modules/drivers depend on this syncq flow control, we prefer to
221  * choose a large number as the default value. For potential
222  * performance gain, this value is tunable in /etc/system.
223  */
224 int sq_max_size = 10000;
225 
226 /*
227  * The number of ciputctrl structures per syncq and stream we create when
228  * needed.
229  */
230 int n_ciputctrl;
231 int max_n_ciputctrl = 16;
232 /*
233  * If n_ciputctrl is < min_n_ciputctrl don't even create ciputctrl_cache.
234  */
235 int min_n_ciputctrl = 2;
236 
237 /*
238  * Per-driver/module syncqs
239  * ========================
240  *
241  * For drivers/modules that use PERMOD or outer syncqs we keep a list of
242  * perdm structures, new entries being added (and new syncqs allocated) when
243  * setq() encounters a module/driver with a streamtab that it hasn't seen
244  * before.
245  * The reason for this mechanism is that some modules and drivers share a
246  * common streamtab and it is necessary for those modules and drivers to also
247  * share a common PERMOD syncq.
248  *
249  * perdm_list --> dm_str == streamtab_1
250  *                dm_sq == syncq_1
251  *                dm_ref
252  *                dm_next --> dm_str == streamtab_2
253  *                            dm_sq == syncq_2
254  *                            dm_ref
255  *                            dm_next --> ... NULL
256  *
257  * The dm_ref field is incremented for each new driver/module that takes
258  * a reference to the perdm structure and hence shares the syncq.
259  * References are held in the fmodsw_impl_t structure for each STREAMS module
260  * or the dev_impl array (indexed by device major number) for each driver.
261  *
262  * perdm_list -> [dm_ref == 1] -> [dm_ref == 2] -> [dm_ref == 1] -> NULL
263  *		     ^                 ^ ^               ^
264  *                   |  ______________/  |               |
265  *                   | /                 |               |
266  * dev_impl:     ...|x|y|...          module A	      module B
267  *
268  * When a module/driver is unloaded the reference count is decremented and,
269  * when it falls to zero, the perdm structure is removed from the list and
270  * the syncq is freed (see rele_dm()).
271  */
272 perdm_t *perdm_list = NULL;
273 static krwlock_t perdm_rwlock;
274 cdevsw_impl_t *devimpl;
275 
276 extern struct qinit strdata;
277 extern struct qinit stwdata;
278 
279 static void runservice(queue_t *);
280 static void streams_bufcall_service(void);
281 static void streams_qbkgrnd_service(void);
282 static void streams_sqbkgrnd_service(void);
283 static syncq_t *new_syncq(void);
284 static void free_syncq(syncq_t *);
285 static void outer_insert(syncq_t *, syncq_t *);
286 static void outer_remove(syncq_t *, syncq_t *);
287 static void write_now(syncq_t *);
288 static void clr_qfull(queue_t *);
289 static void runbufcalls(void);
290 static void sqenable(syncq_t *);
291 static void sqfill_events(syncq_t *, queue_t *, mblk_t *, void (*)());
292 static void wait_q_syncq(queue_t *);
293 static void backenable_insertedq(queue_t *);
294 
295 static void queue_service(queue_t *);
296 static void stream_service(stdata_t *);
297 static void syncq_service(syncq_t *);
298 static void qwriter_outer_service(syncq_t *);
299 static void mblk_free(mblk_t *);
300 #ifdef DEBUG
301 static int qprocsareon(queue_t *);
302 #endif
303 
304 static void set_nfsrv_ptr(queue_t *, queue_t *, queue_t *, queue_t *);
305 static void reset_nfsrv_ptr(queue_t *, queue_t *);
306 void set_qfull(queue_t *);
307 
308 static void sq_run_events(syncq_t *);
309 static int propagate_syncq(queue_t *);
310 
311 static void	blocksq(syncq_t *, ushort_t, int);
312 static void	unblocksq(syncq_t *, ushort_t, int);
313 static int	dropsq(syncq_t *, uint16_t);
314 static void	emptysq(syncq_t *);
315 static sqlist_t *sqlist_alloc(struct stdata *, int);
316 static void	sqlist_free(sqlist_t *);
317 static sqlist_t	*sqlist_build(queue_t *, struct stdata *, boolean_t);
318 static void	sqlist_insert(sqlist_t *, syncq_t *);
319 static void	sqlist_insertall(sqlist_t *, queue_t *);
320 
321 static void	strsetuio(stdata_t *);
322 
323 struct kmem_cache *stream_head_cache;
324 struct kmem_cache *queue_cache;
325 struct kmem_cache *syncq_cache;
326 struct kmem_cache *qband_cache;
327 struct kmem_cache *linkinfo_cache;
328 struct kmem_cache *ciputctrl_cache = NULL;
329 
330 static linkinfo_t *linkinfo_list;
331 
332 /* Global esballoc throttling queue */
333 static esb_queue_t system_esbq;
334 
335 /* Array of esballoc throttling queues, of length esbq_nelem */
336 static esb_queue_t *volatile system_esbq_array;
337 static int esbq_nelem;
338 static kmutex_t esbq_lock;
339 static int esbq_log2_cpus_per_q = 0;
340 
341 /* Scale the system_esbq length by setting number of CPUs per queue. */
342 uint_t esbq_cpus_per_q = 1;
343 
344 /*
345  * esballoc tunable parameters.
346  */
347 int		esbq_max_qlen = 0x16;	/* throttled queue length */
348 clock_t		esbq_timeout = 0x8;	/* timeout to process esb queue */
349 
350 /*
351  * Routines to handle esballoc queueing.
352  */
353 static void esballoc_process_queue(esb_queue_t *);
354 static void esballoc_enqueue_mblk(mblk_t *);
355 static void esballoc_timer(void *);
356 static void esballoc_set_timer(esb_queue_t *, clock_t);
357 static void esballoc_mblk_free(mblk_t *);
358 
359 /*
360  *  Qinit structure and Module_info structures
361  *	for passthru read and write queues
362  */
363 
364 static int pass_rput(queue_t *, mblk_t *);
365 static int pass_wput(queue_t *, mblk_t *);
366 static queue_t *link_addpassthru(stdata_t *);
367 static void link_rempassthru(queue_t *);
368 
369 struct  module_info passthru_info = {
370 	0,
371 	"passthru",
372 	0,
373 	INFPSZ,
374 	STRHIGH,
375 	STRLOW
376 };
377 
378 struct  qinit passthru_rinit = {
379 	pass_rput,
380 	NULL,
381 	NULL,
382 	NULL,
383 	NULL,
384 	&passthru_info,
385 	NULL
386 };
387 
388 struct  qinit passthru_winit = {
389 	pass_wput,
390 	NULL,
391 	NULL,
392 	NULL,
393 	NULL,
394 	&passthru_info,
395 	NULL
396 };
397 
398 /*
399  * Verify correctness of list head/tail pointers.
400  */
401 #define	LISTCHECK(head, tail, link) {				\
402 	EQUIV(head, tail);					\
403 	IMPLY(tail != NULL, tail->link == NULL);		\
404 }
405 
406 /*
407  * Enqueue a list element `el' in the end of a list denoted by `head' and `tail'
408  * using a `link' field.
409  */
410 #define	ENQUEUE(el, head, tail, link) {				\
411 	ASSERT(el->link == NULL);				\
412 	LISTCHECK(head, tail, link);				\
413 	if (head == NULL)					\
414 		head = el;					\
415 	else							\
416 		tail->link = el;				\
417 	tail = el;						\
418 }
419 
420 /*
421  * Dequeue the first element of the list denoted by `head' and `tail' pointers
422  * using a `link' field and put result into `el'.
423  */
424 #define	DQ(el, head, tail, link) {				\
425 	LISTCHECK(head, tail, link);				\
426 	el = head;						\
427 	if (head != NULL) {					\
428 		head = head->link;				\
429 		if (head == NULL)				\
430 			tail = NULL;				\
431 		el->link = NULL;				\
432 	}							\
433 }
434 
435 /*
436  * Remove `el' from the list using `chase' and `curr' pointers and return result
437  * in `succeed'.
438  */
439 #define	RMQ(el, head, tail, link, chase, curr, succeed) {	\
440 	LISTCHECK(head, tail, link);				\
441 	chase = NULL;						\
442 	succeed = 0;						\
443 	for (curr = head; (curr != el) && (curr != NULL); curr = curr->link) \
444 		chase = curr;					\
445 	if (curr != NULL) {					\
446 		succeed = 1;					\
447 		ASSERT(curr == el);				\
448 		if (chase != NULL)				\
449 			chase->link = curr->link;		\
450 		else						\
451 			head = curr->link;			\
452 		curr->link = NULL;				\
453 		if (curr == tail)				\
454 			tail = chase;				\
455 	}							\
456 	LISTCHECK(head, tail, link);				\
457 }
458 
459 /* Handling of delayed messages on the inner syncq. */
460 
461 /*
462  * DEBUG versions should use function versions (to simplify tracing) and
463  * non-DEBUG kernels should use macro versions.
464  */
465 
466 /*
467  * Put a queue on the syncq list of queues.
468  * Assumes SQLOCK held.
469  */
470 #define	SQPUT_Q(sq, qp)							\
471 {									\
472 	ASSERT(MUTEX_HELD(SQLOCK(sq)));					\
473 	if (!(qp->q_sqflags & Q_SQQUEUED)) {				\
474 		/* The queue should not be linked anywhere */		\
475 		ASSERT((qp->q_sqprev == NULL) && (qp->q_sqnext == NULL)); \
476 		/* Head and tail may only be NULL simultaneously */	\
477 		EQUIV(sq->sq_head, sq->sq_tail);			\
478 		/* Queue may be only enqueued on its syncq */		\
479 		ASSERT(sq == qp->q_syncq);				\
480 		/* Check the correctness of SQ_MESSAGES flag */		\
481 		EQUIV(sq->sq_head, (sq->sq_flags & SQ_MESSAGES));	\
482 		/* Sanity check first/last elements of the list */	\
483 		IMPLY(sq->sq_head != NULL, sq->sq_head->q_sqprev == NULL);\
484 		IMPLY(sq->sq_tail != NULL, sq->sq_tail->q_sqnext == NULL);\
485 		/*							\
486 		 * Sanity check of priority field: empty queue should	\
487 		 * have zero priority					\
488 		 * and nqueues equal to zero.				\
489 		 */							\
490 		IMPLY(sq->sq_head == NULL, sq->sq_pri == 0);		\
491 		/* Sanity check of sq_nqueues field */			\
492 		EQUIV(sq->sq_head, sq->sq_nqueues);			\
493 		if (sq->sq_head == NULL) {				\
494 			sq->sq_head = sq->sq_tail = qp;			\
495 			sq->sq_flags |= SQ_MESSAGES;			\
496 		} else if (qp->q_spri == 0) {				\
497 			qp->q_sqprev = sq->sq_tail;			\
498 			sq->sq_tail->q_sqnext = qp;			\
499 			sq->sq_tail = qp;				\
500 		} else {						\
501 			/*						\
502 			 * Put this queue in priority order: higher	\
503 			 * priority gets closer to the head.		\
504 			 */						\
505 			queue_t **qpp = &sq->sq_tail;			\
506 			queue_t *qnext = NULL;				\
507 									\
508 			while (*qpp != NULL && qp->q_spri > (*qpp)->q_spri) { \
509 				qnext = *qpp;				\
510 				qpp = &(*qpp)->q_sqprev;		\
511 			}						\
512 			qp->q_sqnext = qnext;				\
513 			qp->q_sqprev = *qpp;				\
514 			if (*qpp != NULL) {				\
515 				(*qpp)->q_sqnext = qp;			\
516 			} else {					\
517 				sq->sq_head = qp;			\
518 				sq->sq_pri = sq->sq_head->q_spri;	\
519 			}						\
520 			*qpp = qp;					\
521 		}							\
522 		qp->q_sqflags |= Q_SQQUEUED;				\
523 		qp->q_sqtstamp = ddi_get_lbolt();			\
524 		sq->sq_nqueues++;					\
525 	}								\
526 }
527 
528 /*
529  * Remove a queue from the syncq list
530  * Assumes SQLOCK held.
531  */
532 #define	SQRM_Q(sq, qp)							\
533 	{								\
534 		ASSERT(MUTEX_HELD(SQLOCK(sq)));				\
535 		ASSERT(qp->q_sqflags & Q_SQQUEUED);			\
536 		ASSERT(sq->sq_head != NULL && sq->sq_tail != NULL);	\
537 		ASSERT((sq->sq_flags & SQ_MESSAGES) != 0);		\
538 		/* Check that the queue is actually in the list */	\
539 		ASSERT(qp->q_sqnext != NULL || sq->sq_tail == qp);	\
540 		ASSERT(qp->q_sqprev != NULL || sq->sq_head == qp);	\
541 		ASSERT(sq->sq_nqueues != 0);				\
542 		if (qp->q_sqprev == NULL) {				\
543 			/* First queue on list, make head q_sqnext */	\
544 			sq->sq_head = qp->q_sqnext;			\
545 		} else {						\
546 			/* Make prev->next == next */			\
547 			qp->q_sqprev->q_sqnext = qp->q_sqnext;		\
548 		}							\
549 		if (qp->q_sqnext == NULL) {				\
550 			/* Last queue on list, make tail sqprev */	\
551 			sq->sq_tail = qp->q_sqprev;			\
552 		} else {						\
553 			/* Make next->prev == prev */			\
554 			qp->q_sqnext->q_sqprev = qp->q_sqprev;		\
555 		}							\
556 		/* clear out references on this queue */		\
557 		qp->q_sqprev = qp->q_sqnext = NULL;			\
558 		qp->q_sqflags &= ~Q_SQQUEUED;				\
559 		/* If there is nothing queued, clear SQ_MESSAGES */	\
560 		if (sq->sq_head != NULL) {				\
561 			sq->sq_pri = sq->sq_head->q_spri;		\
562 		} else	{						\
563 			sq->sq_flags &= ~SQ_MESSAGES;			\
564 			sq->sq_pri = 0;					\
565 		}							\
566 		sq->sq_nqueues--;					\
567 		ASSERT(sq->sq_head != NULL || sq->sq_evhead != NULL ||	\
568 		    (sq->sq_flags & SQ_QUEUED) == 0);			\
569 	}
570 
571 /* Hide the definition from the header file. */
572 #ifdef SQPUT_MP
573 #undef SQPUT_MP
574 #endif
575 
576 /*
577  * Put a message on the queue syncq.
578  * Assumes QLOCK held.
579  */
580 #define	SQPUT_MP(qp, mp)						\
581 	{								\
582 		ASSERT(MUTEX_HELD(QLOCK(qp)));				\
583 		ASSERT(qp->q_sqhead == NULL ||				\
584 		    (qp->q_sqtail != NULL &&				\
585 		    qp->q_sqtail->b_next == NULL));			\
586 		qp->q_syncqmsgs++;					\
587 		ASSERT(qp->q_syncqmsgs != 0);	/* Wraparound */	\
588 		if (qp->q_sqhead == NULL) {				\
589 			qp->q_sqhead = qp->q_sqtail = mp;		\
590 		} else {						\
591 			qp->q_sqtail->b_next = mp;			\
592 			qp->q_sqtail = mp;				\
593 		}							\
594 		ASSERT(qp->q_syncqmsgs > 0);				\
595 		set_qfull(qp);						\
596 	}
597 
598 #define	SQ_PUTCOUNT_SETFAST_LOCKED(sq) {				\
599 		ASSERT(MUTEX_HELD(SQLOCK(sq)));				\
600 		if ((sq)->sq_ciputctrl != NULL) {			\
601 			int i;						\
602 			int nlocks = (sq)->sq_nciputctrl;		\
603 			ciputctrl_t *cip = (sq)->sq_ciputctrl;		\
604 			ASSERT((sq)->sq_type & SQ_CIPUT);		\
605 			for (i = 0; i <= nlocks; i++) {			\
606 				ASSERT(MUTEX_HELD(&cip[i].ciputctrl_lock)); \
607 				cip[i].ciputctrl_count |= SQ_FASTPUT;	\
608 			}						\
609 		}							\
610 	}
611 
612 
613 #define	SQ_PUTCOUNT_CLRFAST_LOCKED(sq) {				\
614 		ASSERT(MUTEX_HELD(SQLOCK(sq)));				\
615 		if ((sq)->sq_ciputctrl != NULL) {			\
616 			int i;						\
617 			int nlocks = (sq)->sq_nciputctrl;		\
618 			ciputctrl_t *cip = (sq)->sq_ciputctrl;		\
619 			ASSERT((sq)->sq_type & SQ_CIPUT);		\
620 			for (i = 0; i <= nlocks; i++) {			\
621 				ASSERT(MUTEX_HELD(&cip[i].ciputctrl_lock)); \
622 				cip[i].ciputctrl_count &= ~SQ_FASTPUT;	\
623 			}						\
624 		}							\
625 	}
626 
627 /*
628  * Run service procedures for all queues in the stream head.
629  */
630 #define	STR_SERVICE(stp, q) {						\
631 	ASSERT(MUTEX_HELD(&stp->sd_qlock));				\
632 	while (stp->sd_qhead != NULL) {					\
633 		DQ(q, stp->sd_qhead, stp->sd_qtail, q_link);		\
634 		ASSERT(stp->sd_nqueues > 0);				\
635 		stp->sd_nqueues--;					\
636 		ASSERT(!(q->q_flag & QINSERVICE));			\
637 		mutex_exit(&stp->sd_qlock);				\
638 		queue_service(q);					\
639 		mutex_enter(&stp->sd_qlock);				\
640 	}								\
641 	ASSERT(stp->sd_nqueues == 0);					\
642 	ASSERT((stp->sd_qhead == NULL) && (stp->sd_qtail == NULL));	\
643 }
644 
645 /*
646  * Constructor/destructor routines for the stream head cache
647  */
648 /* ARGSUSED */
649 static int
stream_head_constructor(void * buf,void * cdrarg,int kmflags)650 stream_head_constructor(void *buf, void *cdrarg, int kmflags)
651 {
652 	stdata_t *stp = buf;
653 
654 	mutex_init(&stp->sd_lock, NULL, MUTEX_DEFAULT, NULL);
655 	mutex_init(&stp->sd_reflock, NULL, MUTEX_DEFAULT, NULL);
656 	mutex_init(&stp->sd_qlock, NULL, MUTEX_DEFAULT, NULL);
657 	cv_init(&stp->sd_monitor, NULL, CV_DEFAULT, NULL);
658 	cv_init(&stp->sd_iocmonitor, NULL, CV_DEFAULT, NULL);
659 	cv_init(&stp->sd_refmonitor, NULL, CV_DEFAULT, NULL);
660 	cv_init(&stp->sd_qcv, NULL, CV_DEFAULT, NULL);
661 	cv_init(&stp->sd_zcopy_wait, NULL, CV_DEFAULT, NULL);
662 	stp->sd_wrq = NULL;
663 
664 	return (0);
665 }
666 
667 /* ARGSUSED */
668 static void
stream_head_destructor(void * buf,void * cdrarg)669 stream_head_destructor(void *buf, void *cdrarg)
670 {
671 	stdata_t *stp = buf;
672 
673 	mutex_destroy(&stp->sd_lock);
674 	mutex_destroy(&stp->sd_reflock);
675 	mutex_destroy(&stp->sd_qlock);
676 	cv_destroy(&stp->sd_monitor);
677 	cv_destroy(&stp->sd_iocmonitor);
678 	cv_destroy(&stp->sd_refmonitor);
679 	cv_destroy(&stp->sd_qcv);
680 	cv_destroy(&stp->sd_zcopy_wait);
681 }
682 
683 /*
684  * Constructor/destructor routines for the queue cache
685  */
686 /* ARGSUSED */
687 static int
queue_constructor(void * buf,void * cdrarg,int kmflags)688 queue_constructor(void *buf, void *cdrarg, int kmflags)
689 {
690 	queinfo_t *qip = buf;
691 	queue_t *qp = &qip->qu_rqueue;
692 	queue_t *wqp = &qip->qu_wqueue;
693 	syncq_t	*sq = &qip->qu_syncq;
694 
695 	qp->q_first = NULL;
696 	qp->q_link = NULL;
697 	qp->q_count = 0;
698 	qp->q_mblkcnt = 0;
699 	qp->q_sqhead = NULL;
700 	qp->q_sqtail = NULL;
701 	qp->q_sqnext = NULL;
702 	qp->q_sqprev = NULL;
703 	qp->q_sqflags = 0;
704 	qp->q_rwcnt = 0;
705 	qp->q_spri = 0;
706 
707 	mutex_init(QLOCK(qp), NULL, MUTEX_DEFAULT, NULL);
708 	cv_init(&qp->q_wait, NULL, CV_DEFAULT, NULL);
709 
710 	wqp->q_first = NULL;
711 	wqp->q_link = NULL;
712 	wqp->q_count = 0;
713 	wqp->q_mblkcnt = 0;
714 	wqp->q_sqhead = NULL;
715 	wqp->q_sqtail = NULL;
716 	wqp->q_sqnext = NULL;
717 	wqp->q_sqprev = NULL;
718 	wqp->q_sqflags = 0;
719 	wqp->q_rwcnt = 0;
720 	wqp->q_spri = 0;
721 
722 	mutex_init(QLOCK(wqp), NULL, MUTEX_DEFAULT, NULL);
723 	cv_init(&wqp->q_wait, NULL, CV_DEFAULT, NULL);
724 
725 	sq->sq_head = NULL;
726 	sq->sq_tail = NULL;
727 	sq->sq_evhead = NULL;
728 	sq->sq_evtail = NULL;
729 	sq->sq_callbpend = NULL;
730 	sq->sq_outer = NULL;
731 	sq->sq_onext = NULL;
732 	sq->sq_oprev = NULL;
733 	sq->sq_next = NULL;
734 	sq->sq_svcflags = 0;
735 	sq->sq_servcount = 0;
736 	sq->sq_needexcl = 0;
737 	sq->sq_nqueues = 0;
738 	sq->sq_pri = 0;
739 
740 	mutex_init(&sq->sq_lock, NULL, MUTEX_DEFAULT, NULL);
741 	cv_init(&sq->sq_wait, NULL, CV_DEFAULT, NULL);
742 	cv_init(&sq->sq_exitwait, NULL, CV_DEFAULT, NULL);
743 
744 	return (0);
745 }
746 
747 /* ARGSUSED */
748 static void
queue_destructor(void * buf,void * cdrarg)749 queue_destructor(void *buf, void *cdrarg)
750 {
751 	queinfo_t *qip = buf;
752 	queue_t *qp = &qip->qu_rqueue;
753 	queue_t *wqp = &qip->qu_wqueue;
754 	syncq_t	*sq = &qip->qu_syncq;
755 
756 	ASSERT(qp->q_sqhead == NULL);
757 	ASSERT(wqp->q_sqhead == NULL);
758 	ASSERT(qp->q_sqnext == NULL);
759 	ASSERT(wqp->q_sqnext == NULL);
760 	ASSERT(qp->q_rwcnt == 0);
761 	ASSERT(wqp->q_rwcnt == 0);
762 
763 	mutex_destroy(&qp->q_lock);
764 	cv_destroy(&qp->q_wait);
765 
766 	mutex_destroy(&wqp->q_lock);
767 	cv_destroy(&wqp->q_wait);
768 
769 	mutex_destroy(&sq->sq_lock);
770 	cv_destroy(&sq->sq_wait);
771 	cv_destroy(&sq->sq_exitwait);
772 }
773 
774 /*
775  * Constructor/destructor routines for the syncq cache
776  */
777 /* ARGSUSED */
778 static int
syncq_constructor(void * buf,void * cdrarg,int kmflags)779 syncq_constructor(void *buf, void *cdrarg, int kmflags)
780 {
781 	syncq_t	*sq = buf;
782 
783 	bzero(buf, sizeof (syncq_t));
784 
785 	mutex_init(&sq->sq_lock, NULL, MUTEX_DEFAULT, NULL);
786 	cv_init(&sq->sq_wait, NULL, CV_DEFAULT, NULL);
787 	cv_init(&sq->sq_exitwait, NULL, CV_DEFAULT, NULL);
788 
789 	return (0);
790 }
791 
792 /* ARGSUSED */
793 static void
syncq_destructor(void * buf,void * cdrarg)794 syncq_destructor(void *buf, void *cdrarg)
795 {
796 	syncq_t	*sq = buf;
797 
798 	ASSERT(sq->sq_head == NULL);
799 	ASSERT(sq->sq_tail == NULL);
800 	ASSERT(sq->sq_evhead == NULL);
801 	ASSERT(sq->sq_evtail == NULL);
802 	ASSERT(sq->sq_callbpend == NULL);
803 	ASSERT(sq->sq_callbflags == 0);
804 	ASSERT(sq->sq_outer == NULL);
805 	ASSERT(sq->sq_onext == NULL);
806 	ASSERT(sq->sq_oprev == NULL);
807 	ASSERT(sq->sq_next == NULL);
808 	ASSERT(sq->sq_needexcl == 0);
809 	ASSERT(sq->sq_svcflags == 0);
810 	ASSERT(sq->sq_servcount == 0);
811 	ASSERT(sq->sq_nqueues == 0);
812 	ASSERT(sq->sq_pri == 0);
813 	ASSERT(sq->sq_count == 0);
814 	ASSERT(sq->sq_rmqcount == 0);
815 	ASSERT(sq->sq_cancelid == 0);
816 	ASSERT(sq->sq_ciputctrl == NULL);
817 	ASSERT(sq->sq_nciputctrl == 0);
818 	ASSERT(sq->sq_type == 0);
819 	ASSERT(sq->sq_flags == 0);
820 
821 	mutex_destroy(&sq->sq_lock);
822 	cv_destroy(&sq->sq_wait);
823 	cv_destroy(&sq->sq_exitwait);
824 }
825 
826 /* ARGSUSED */
827 static int
ciputctrl_constructor(void * buf,void * cdrarg,int kmflags)828 ciputctrl_constructor(void *buf, void *cdrarg, int kmflags)
829 {
830 	ciputctrl_t *cip = buf;
831 	int i;
832 
833 	for (i = 0; i < n_ciputctrl; i++) {
834 		cip[i].ciputctrl_count = SQ_FASTPUT;
835 		mutex_init(&cip[i].ciputctrl_lock, NULL, MUTEX_DEFAULT, NULL);
836 	}
837 
838 	return (0);
839 }
840 
841 /* ARGSUSED */
842 static void
ciputctrl_destructor(void * buf,void * cdrarg)843 ciputctrl_destructor(void *buf, void *cdrarg)
844 {
845 	ciputctrl_t *cip = buf;
846 	int i;
847 
848 	for (i = 0; i < n_ciputctrl; i++) {
849 		ASSERT(cip[i].ciputctrl_count & SQ_FASTPUT);
850 		mutex_destroy(&cip[i].ciputctrl_lock);
851 	}
852 }
853 
854 /*
855  * Init routine run from main at boot time.
856  */
857 void
strinit(void)858 strinit(void)
859 {
860 	int ncpus = ((boot_max_ncpus == -1) ? max_ncpus : boot_max_ncpus);
861 
862 	stream_head_cache = kmem_cache_create("stream_head_cache",
863 	    sizeof (stdata_t), 0,
864 	    stream_head_constructor, stream_head_destructor, NULL,
865 	    NULL, NULL, 0);
866 
867 	queue_cache = kmem_cache_create("queue_cache", sizeof (queinfo_t), 0,
868 	    queue_constructor, queue_destructor, NULL, NULL, NULL, 0);
869 
870 	syncq_cache = kmem_cache_create("syncq_cache", sizeof (syncq_t), 0,
871 	    syncq_constructor, syncq_destructor, NULL, NULL, NULL, 0);
872 
873 	qband_cache = kmem_cache_create("qband_cache",
874 	    sizeof (qband_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
875 
876 	linkinfo_cache = kmem_cache_create("linkinfo_cache",
877 	    sizeof (linkinfo_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
878 
879 	n_ciputctrl = ncpus;
880 	n_ciputctrl = 1 << highbit(n_ciputctrl - 1);
881 	ASSERT(n_ciputctrl >= 1);
882 	n_ciputctrl = MIN(n_ciputctrl, max_n_ciputctrl);
883 	if (n_ciputctrl >= min_n_ciputctrl) {
884 		ciputctrl_cache = kmem_cache_create("ciputctrl_cache",
885 		    sizeof (ciputctrl_t) * n_ciputctrl,
886 		    sizeof (ciputctrl_t), ciputctrl_constructor,
887 		    ciputctrl_destructor, NULL, NULL, NULL, 0);
888 	}
889 
890 	streams_taskq = system_taskq;
891 
892 	if (streams_taskq == NULL)
893 		panic("strinit: no memory for streams taskq!");
894 
895 	bc_bkgrnd_thread = thread_create(NULL, 0,
896 	    streams_bufcall_service, NULL, 0, &p0, TS_RUN, streams_lopri);
897 
898 	streams_qbkgrnd_thread = thread_create(NULL, 0,
899 	    streams_qbkgrnd_service, NULL, 0, &p0, TS_RUN, streams_lopri);
900 
901 	streams_sqbkgrnd_thread = thread_create(NULL, 0,
902 	    streams_sqbkgrnd_service, NULL, 0, &p0, TS_RUN, streams_lopri);
903 
904 	/*
905 	 * Create STREAMS kstats.
906 	 */
907 	str_kstat = kstat_create("streams", 0, "strstat",
908 	    "net", KSTAT_TYPE_NAMED,
909 	    sizeof (str_statistics) / sizeof (kstat_named_t),
910 	    KSTAT_FLAG_VIRTUAL);
911 
912 	if (str_kstat != NULL) {
913 		str_kstat->ks_data = &str_statistics;
914 		kstat_install(str_kstat);
915 	}
916 
917 	/*
918 	 * TPI support routine initialisation.
919 	 */
920 	tpi_init();
921 
922 	/*
923 	 * Handle to have autopush and persistent link information per
924 	 * zone.
925 	 * Note: uses shutdown hook instead of destroy hook so that the
926 	 * persistent links can be torn down before the destroy hooks
927 	 * in the TCP/IP stack are called.
928 	 */
929 	netstack_register(NS_STR, str_stack_init, str_stack_shutdown,
930 	    str_stack_fini);
931 }
932 
933 void
str_sendsig(vnode_t * vp,int event,uchar_t band,int error)934 str_sendsig(vnode_t *vp, int event, uchar_t band, int error)
935 {
936 	struct stdata *stp;
937 
938 	ASSERT(vp->v_stream);
939 	stp = vp->v_stream;
940 	/* Have to hold sd_lock to prevent siglist from changing */
941 	mutex_enter(&stp->sd_lock);
942 	if (stp->sd_sigflags & event)
943 		strsendsig(stp->sd_siglist, event, band, error);
944 	mutex_exit(&stp->sd_lock);
945 }
946 
947 /*
948  * Send the "sevent" set of signals to a process.
949  * This might send more than one signal if the process is registered
950  * for multiple events. The caller should pass in an sevent that only
951  * includes the events for which the process has registered.
952  */
953 static void
dosendsig(proc_t * proc,int events,int sevent,k_siginfo_t * info,uchar_t band,int error)954 dosendsig(proc_t *proc, int events, int sevent, k_siginfo_t *info,
955     uchar_t band, int error)
956 {
957 	ASSERT(MUTEX_HELD(&proc->p_lock));
958 
959 	info->si_band = 0;
960 	info->si_errno = 0;
961 
962 	if (sevent & S_ERROR) {
963 		sevent &= ~S_ERROR;
964 		info->si_code = POLL_ERR;
965 		info->si_errno = error;
966 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
967 		    "strsendsig:proc %p info %p", proc, info);
968 		sigaddq(proc, NULL, info, KM_NOSLEEP);
969 		info->si_errno = 0;
970 	}
971 	if (sevent & S_HANGUP) {
972 		sevent &= ~S_HANGUP;
973 		info->si_code = POLL_HUP;
974 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
975 		    "strsendsig:proc %p info %p", proc, info);
976 		sigaddq(proc, NULL, info, KM_NOSLEEP);
977 	}
978 	if (sevent & S_HIPRI) {
979 		sevent &= ~S_HIPRI;
980 		info->si_code = POLL_PRI;
981 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
982 		    "strsendsig:proc %p info %p", proc, info);
983 		sigaddq(proc, NULL, info, KM_NOSLEEP);
984 	}
985 	if (sevent & S_RDBAND) {
986 		sevent &= ~S_RDBAND;
987 		if (events & S_BANDURG)
988 			sigtoproc(proc, NULL, SIGURG);
989 		else
990 			sigtoproc(proc, NULL, SIGPOLL);
991 	}
992 	if (sevent & S_WRBAND) {
993 		sevent &= ~S_WRBAND;
994 		sigtoproc(proc, NULL, SIGPOLL);
995 	}
996 	if (sevent & S_INPUT) {
997 		sevent &= ~S_INPUT;
998 		info->si_code = POLL_IN;
999 		info->si_band = band;
1000 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
1001 		    "strsendsig:proc %p info %p", proc, info);
1002 		sigaddq(proc, NULL, info, KM_NOSLEEP);
1003 		info->si_band = 0;
1004 	}
1005 	if (sevent & S_OUTPUT) {
1006 		sevent &= ~S_OUTPUT;
1007 		info->si_code = POLL_OUT;
1008 		info->si_band = band;
1009 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
1010 		    "strsendsig:proc %p info %p", proc, info);
1011 		sigaddq(proc, NULL, info, KM_NOSLEEP);
1012 		info->si_band = 0;
1013 	}
1014 	if (sevent & S_MSG) {
1015 		sevent &= ~S_MSG;
1016 		info->si_code = POLL_MSG;
1017 		info->si_band = band;
1018 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
1019 		    "strsendsig:proc %p info %p", proc, info);
1020 		sigaddq(proc, NULL, info, KM_NOSLEEP);
1021 		info->si_band = 0;
1022 	}
1023 	if (sevent & S_RDNORM) {
1024 		sevent &= ~S_RDNORM;
1025 		sigtoproc(proc, NULL, SIGPOLL);
1026 	}
1027 	if (sevent != 0) {
1028 		panic("strsendsig: unknown event(s) %x", sevent);
1029 	}
1030 }
1031 
1032 /*
1033  * Send SIGPOLL/SIGURG signal to all processes and process groups
1034  * registered on the given signal list that want a signal for at
1035  * least one of the specified events.
1036  *
1037  * Must be called with exclusive access to siglist (caller holding sd_lock).
1038  *
1039  * strioctl(I_SETSIG/I_ESETSIG) will only change siglist when holding
1040  * sd_lock and the ioctl code maintains a PID_HOLD on the pid structure
1041  * while it is in the siglist.
1042  *
1043  * For performance reasons (MP scalability) the code drops pidlock
1044  * when sending signals to a single process.
1045  * When sending to a process group the code holds
1046  * pidlock to prevent the membership in the process group from changing
1047  * while walking the p_pglink list.
1048  */
1049 void
strsendsig(strsig_t * siglist,int event,uchar_t band,int error)1050 strsendsig(strsig_t *siglist, int event, uchar_t band, int error)
1051 {
1052 	strsig_t *ssp;
1053 	k_siginfo_t info;
1054 	struct pid *pidp;
1055 	proc_t  *proc;
1056 
1057 	info.si_signo = SIGPOLL;
1058 	info.si_errno = 0;
1059 	for (ssp = siglist; ssp; ssp = ssp->ss_next) {
1060 		int sevent;
1061 
1062 		sevent = ssp->ss_events & event;
1063 		if (sevent == 0)
1064 			continue;
1065 
1066 		if ((pidp = ssp->ss_pidp) == NULL) {
1067 			/* pid was released but still on event list */
1068 			continue;
1069 		}
1070 
1071 
1072 		if (ssp->ss_pid > 0) {
1073 			/*
1074 			 * XXX This unfortunately still generates
1075 			 * a signal when a fd is closed but
1076 			 * the proc is active.
1077 			 */
1078 			ASSERT(ssp->ss_pid == pidp->pid_id);
1079 
1080 			mutex_enter(&pidlock);
1081 			proc = prfind_zone(pidp->pid_id, ALL_ZONES);
1082 			if (proc == NULL) {
1083 				mutex_exit(&pidlock);
1084 				continue;
1085 			}
1086 			mutex_enter(&proc->p_lock);
1087 			mutex_exit(&pidlock);
1088 			dosendsig(proc, ssp->ss_events, sevent, &info,
1089 			    band, error);
1090 			mutex_exit(&proc->p_lock);
1091 		} else {
1092 			/*
1093 			 * Send to process group. Hold pidlock across
1094 			 * calls to dosendsig().
1095 			 */
1096 			pid_t pgrp = -ssp->ss_pid;
1097 
1098 			mutex_enter(&pidlock);
1099 			proc = pgfind_zone(pgrp, ALL_ZONES);
1100 			while (proc != NULL) {
1101 				mutex_enter(&proc->p_lock);
1102 				dosendsig(proc, ssp->ss_events, sevent,
1103 				    &info, band, error);
1104 				mutex_exit(&proc->p_lock);
1105 				proc = proc->p_pglink;
1106 			}
1107 			mutex_exit(&pidlock);
1108 		}
1109 	}
1110 }
1111 
1112 /*
1113  * Attach a stream device or module.
1114  * qp is a read queue; the new queue goes in so its next
1115  * read ptr is the argument, and the write queue corresponding
1116  * to the argument points to this queue. Return 0 on success,
1117  * or a non-zero errno on failure.
1118  */
1119 int
qattach(queue_t * qp,dev_t * devp,int oflag,cred_t * crp,fmodsw_impl_t * fp,boolean_t is_insert)1120 qattach(queue_t *qp, dev_t *devp, int oflag, cred_t *crp, fmodsw_impl_t *fp,
1121     boolean_t is_insert)
1122 {
1123 	major_t			major;
1124 	cdevsw_impl_t		*dp;
1125 	struct streamtab	*str;
1126 	queue_t			*rq;
1127 	queue_t			*wrq;
1128 	uint32_t		qflag;
1129 	uint32_t		sqtype;
1130 	perdm_t			*dmp;
1131 	int			error;
1132 	int			sflag;
1133 
1134 	rq = allocq();
1135 	wrq = _WR(rq);
1136 	STREAM(rq) = STREAM(wrq) = STREAM(qp);
1137 
1138 	if (fp != NULL) {
1139 		str = fp->f_str;
1140 		qflag = fp->f_qflag;
1141 		sqtype = fp->f_sqtype;
1142 		dmp = fp->f_dmp;
1143 		IMPLY((qflag & (QPERMOD | QMTOUTPERIM)), dmp != NULL);
1144 		sflag = MODOPEN;
1145 
1146 		/*
1147 		 * stash away a pointer to the module structure so we can
1148 		 * unref it in qdetach.
1149 		 */
1150 		rq->q_fp = fp;
1151 	} else {
1152 		ASSERT(!is_insert);
1153 
1154 		major = getmajor(*devp);
1155 		dp = &devimpl[major];
1156 
1157 		str = dp->d_str;
1158 		ASSERT(str == STREAMSTAB(major));
1159 
1160 		qflag = dp->d_qflag;
1161 		ASSERT(qflag & QISDRV);
1162 		sqtype = dp->d_sqtype;
1163 
1164 		/* create perdm_t if needed */
1165 		if (NEED_DM(dp->d_dmp, qflag))
1166 			dp->d_dmp = hold_dm(str, qflag, sqtype);
1167 
1168 		dmp = dp->d_dmp;
1169 		sflag = 0;
1170 	}
1171 
1172 	TRACE_2(TR_FAC_STREAMS_FR, TR_QATTACH_FLAGS,
1173 	    "qattach:qflag == %X(%X)", qflag, *devp);
1174 
1175 	/* setq might sleep in allocator - avoid holding locks. */
1176 	setq(rq, str->st_rdinit, str->st_wrinit, dmp, qflag, sqtype, B_FALSE);
1177 
1178 	/*
1179 	 * Before calling the module's open routine, set up the q_next
1180 	 * pointer for inserting a module in the middle of a stream.
1181 	 *
1182 	 * Note that we can always set _QINSERTING and set up q_next
1183 	 * pointer for both inserting and pushing a module.  Then there
1184 	 * is no need for the is_insert parameter.  In insertq(), called
1185 	 * by qprocson(), assume that q_next of the new module always points
1186 	 * to the correct queue and use it for insertion.  Everything should
1187 	 * work out fine.  But in the first release of _I_INSERT, we
1188 	 * distinguish between inserting and pushing to make sure that
1189 	 * pushing a module follows the same code path as before.
1190 	 */
1191 	if (is_insert) {
1192 		rq->q_flag |= _QINSERTING;
1193 		rq->q_next = qp;
1194 	}
1195 
1196 	/*
1197 	 * If there is an outer perimeter get exclusive access during
1198 	 * the open procedure.  Bump up the reference count on the queue.
1199 	 */
1200 	entersq(rq->q_syncq, SQ_OPENCLOSE);
1201 	error = (*rq->q_qinfo->qi_qopen)(rq, devp, oflag, sflag, crp);
1202 	if (error != 0)
1203 		goto failed;
1204 	leavesq(rq->q_syncq, SQ_OPENCLOSE);
1205 	ASSERT(qprocsareon(rq));
1206 	return (0);
1207 
1208 failed:
1209 	rq->q_flag &= ~_QINSERTING;
1210 	if (backq(wrq) != NULL && backq(wrq)->q_next == wrq)
1211 		qprocsoff(rq);
1212 	leavesq(rq->q_syncq, SQ_OPENCLOSE);
1213 	rq->q_next = wrq->q_next = NULL;
1214 	qdetach(rq, 0, 0, crp, B_FALSE);
1215 	return (error);
1216 }
1217 
1218 /*
1219  * Handle second open of stream. For modules, set the
1220  * last argument to MODOPEN and do not pass any open flags.
1221  * Ignore dummydev since this is not the first open.
1222  */
1223 int
qreopen(queue_t * qp,dev_t * devp,int flag,cred_t * crp)1224 qreopen(queue_t *qp, dev_t *devp, int flag, cred_t *crp)
1225 {
1226 	int	error;
1227 	dev_t dummydev;
1228 	queue_t *wqp = _WR(qp);
1229 
1230 	ASSERT(qp->q_flag & QREADR);
1231 	entersq(qp->q_syncq, SQ_OPENCLOSE);
1232 
1233 	dummydev = *devp;
1234 	if (error = ((*qp->q_qinfo->qi_qopen)(qp, &dummydev,
1235 	    (wqp->q_next ? 0 : flag), (wqp->q_next ? MODOPEN : 0), crp))) {
1236 		leavesq(qp->q_syncq, SQ_OPENCLOSE);
1237 		mutex_enter(&STREAM(qp)->sd_lock);
1238 		qp->q_stream->sd_flag |= STREOPENFAIL;
1239 		mutex_exit(&STREAM(qp)->sd_lock);
1240 		return (error);
1241 	}
1242 	leavesq(qp->q_syncq, SQ_OPENCLOSE);
1243 
1244 	/*
1245 	 * successful open should have done qprocson()
1246 	 */
1247 	ASSERT(qprocsareon(_RD(qp)));
1248 	return (0);
1249 }
1250 
1251 /*
1252  * Detach a stream module or device.
1253  * If clmode == 1 then the module or driver was opened and its
1254  * close routine must be called. If clmode == 0, the module
1255  * or driver was never opened or the open failed, and so its close
1256  * should not be called.
1257  */
1258 void
qdetach(queue_t * qp,int clmode,int flag,cred_t * crp,boolean_t is_remove)1259 qdetach(queue_t *qp, int clmode, int flag, cred_t *crp, boolean_t is_remove)
1260 {
1261 	queue_t *wqp = _WR(qp);
1262 	ASSERT(STREAM(qp)->sd_flag & (STRCLOSE|STWOPEN|STRPLUMB));
1263 
1264 	if (STREAM_NEEDSERVICE(STREAM(qp)))
1265 		stream_runservice(STREAM(qp));
1266 
1267 	if (clmode) {
1268 		/*
1269 		 * Make sure that all the messages on the write side syncq are
1270 		 * processed and nothing is left. Since we are closing, no new
1271 		 * messages may appear there.
1272 		 */
1273 		wait_q_syncq(wqp);
1274 
1275 		entersq(qp->q_syncq, SQ_OPENCLOSE);
1276 		if (is_remove) {
1277 			mutex_enter(QLOCK(qp));
1278 			qp->q_flag |= _QREMOVING;
1279 			mutex_exit(QLOCK(qp));
1280 		}
1281 		(*qp->q_qinfo->qi_qclose)(qp, flag, crp);
1282 		/*
1283 		 * Check that qprocsoff() was actually called.
1284 		 */
1285 		ASSERT((qp->q_flag & QWCLOSE) && (wqp->q_flag & QWCLOSE));
1286 
1287 		leavesq(qp->q_syncq, SQ_OPENCLOSE);
1288 	} else {
1289 		disable_svc(qp);
1290 	}
1291 
1292 	/*
1293 	 * Allow any threads blocked in entersq to proceed and discover
1294 	 * the QWCLOSE is set.
1295 	 * Note: This assumes that all users of entersq check QWCLOSE.
1296 	 * Currently runservice is the only entersq that can happen
1297 	 * after removeq has finished.
1298 	 * Removeq will have discarded all messages destined to the closing
1299 	 * pair of queues from the syncq.
1300 	 * NOTE: Calling a function inside an assert is unconventional.
1301 	 * However, it does not cause any problem since flush_syncq() does
1302 	 * not change any state except when it returns non-zero i.e.
1303 	 * when the assert will trigger.
1304 	 */
1305 	ASSERT(flush_syncq(qp->q_syncq, qp) == 0);
1306 	ASSERT(flush_syncq(wqp->q_syncq, wqp) == 0);
1307 	ASSERT((qp->q_flag & QPERMOD) ||
1308 	    ((qp->q_syncq->sq_head == NULL) &&
1309 	    (wqp->q_syncq->sq_head == NULL)));
1310 
1311 	/* release any fmodsw_impl_t structure held on behalf of the queue */
1312 	ASSERT(qp->q_fp != NULL || qp->q_flag & QISDRV);
1313 	if (qp->q_fp != NULL)
1314 		fmodsw_rele(qp->q_fp);
1315 
1316 	/* freeq removes us from the outer perimeter if any */
1317 	freeq(qp);
1318 }
1319 
1320 /* Prevent service procedures from being called */
1321 void
disable_svc(queue_t * qp)1322 disable_svc(queue_t *qp)
1323 {
1324 	queue_t *wqp = _WR(qp);
1325 
1326 	ASSERT(qp->q_flag & QREADR);
1327 	mutex_enter(QLOCK(qp));
1328 	qp->q_flag |= QWCLOSE;
1329 	mutex_exit(QLOCK(qp));
1330 	mutex_enter(QLOCK(wqp));
1331 	wqp->q_flag |= QWCLOSE;
1332 	mutex_exit(QLOCK(wqp));
1333 }
1334 
1335 /* Allow service procedures to be called again */
1336 void
enable_svc(queue_t * qp)1337 enable_svc(queue_t *qp)
1338 {
1339 	queue_t *wqp = _WR(qp);
1340 
1341 	ASSERT(qp->q_flag & QREADR);
1342 	mutex_enter(QLOCK(qp));
1343 	qp->q_flag &= ~QWCLOSE;
1344 	mutex_exit(QLOCK(qp));
1345 	mutex_enter(QLOCK(wqp));
1346 	wqp->q_flag &= ~QWCLOSE;
1347 	mutex_exit(QLOCK(wqp));
1348 }
1349 
1350 /*
1351  * Remove queue from qhead/qtail if it is enabled.
1352  * Only reset QENAB if the queue was removed from the runlist.
1353  * A queue goes through 3 stages:
1354  *	It is on the service list and QENAB is set.
1355  *	It is removed from the service list but QENAB is still set.
1356  *	QENAB gets changed to QINSERVICE.
1357  *	QINSERVICE is reset (when the service procedure is done)
1358  * Thus we can not reset QENAB unless we actually removed it from the service
1359  * queue.
1360  */
1361 void
remove_runlist(queue_t * qp)1362 remove_runlist(queue_t *qp)
1363 {
1364 	if (qp->q_flag & QENAB && qhead != NULL) {
1365 		queue_t *q_chase;
1366 		queue_t *q_curr;
1367 		int removed;
1368 
1369 		mutex_enter(&service_queue);
1370 		RMQ(qp, qhead, qtail, q_link, q_chase, q_curr, removed);
1371 		mutex_exit(&service_queue);
1372 		if (removed) {
1373 			STRSTAT(qremoved);
1374 			qp->q_flag &= ~QENAB;
1375 		}
1376 	}
1377 }
1378 
1379 
1380 /*
1381  * Wait for any pending service processing to complete.
1382  * The removal of queues from the runlist is not atomic with the
1383  * clearing of the QENABLED flag and setting the INSERVICE flag.
1384  * consequently it is possible for remove_runlist in strclose
1385  * to not find the queue on the runlist but for it to be QENABLED
1386  * and not yet INSERVICE -> hence wait_svc needs to check QENABLED
1387  * as well as INSERVICE.
1388  */
1389 void
wait_svc(queue_t * qp)1390 wait_svc(queue_t *qp)
1391 {
1392 	queue_t *wqp = _WR(qp);
1393 
1394 	ASSERT(qp->q_flag & QREADR);
1395 
1396 	/*
1397 	 * Try to remove queues from qhead/qtail list.
1398 	 */
1399 	if (qhead != NULL) {
1400 		remove_runlist(qp);
1401 		remove_runlist(wqp);
1402 	}
1403 	/*
1404 	 * Wait till the syncqs associated with the queue disappear from the
1405 	 * background processing list.
1406 	 * This only needs to be done for non-PERMOD perimeters since
1407 	 * for PERMOD perimeters the syncq may be shared and will only be freed
1408 	 * when the last module/driver is unloaded.
1409 	 * If for PERMOD perimeters queue was on the syncq list, removeq()
1410 	 * should call propagate_syncq() or drain_syncq() for it. Both of these
1411 	 * functions remove the queue from its syncq list, so sqthread will not
1412 	 * try to access the queue.
1413 	 */
1414 	if (!(qp->q_flag & QPERMOD)) {
1415 		syncq_t *rsq = qp->q_syncq;
1416 		syncq_t *wsq = wqp->q_syncq;
1417 
1418 		/*
1419 		 * Disable rsq and wsq and wait for any background processing of
1420 		 * syncq to complete.
1421 		 */
1422 		wait_sq_svc(rsq);
1423 		if (wsq != rsq)
1424 			wait_sq_svc(wsq);
1425 	}
1426 
1427 	mutex_enter(QLOCK(qp));
1428 	while (qp->q_flag & (QINSERVICE|QENAB))
1429 		cv_wait(&qp->q_wait, QLOCK(qp));
1430 	mutex_exit(QLOCK(qp));
1431 	mutex_enter(QLOCK(wqp));
1432 	while (wqp->q_flag & (QINSERVICE|QENAB))
1433 		cv_wait(&wqp->q_wait, QLOCK(wqp));
1434 	mutex_exit(QLOCK(wqp));
1435 }
1436 
1437 /*
1438  * Put ioctl data from userland buffer `arg' into the mblk chain `bp'.
1439  * `flag' must always contain either K_TO_K or U_TO_K; STR_NOSIG may
1440  * also be set, and is passed through to allocb_cred_wait().
1441  *
1442  * Returns errno on failure, zero on success.
1443  */
1444 int
putiocd(mblk_t * bp,char * arg,int flag,cred_t * cr)1445 putiocd(mblk_t *bp, char *arg, int flag, cred_t *cr)
1446 {
1447 	mblk_t *tmp;
1448 	ssize_t  count;
1449 	int error = 0;
1450 
1451 	ASSERT((flag & (U_TO_K | K_TO_K)) == U_TO_K ||
1452 	    (flag & (U_TO_K | K_TO_K)) == K_TO_K);
1453 
1454 	if (bp->b_datap->db_type == M_IOCTL) {
1455 		count = ((struct iocblk *)bp->b_rptr)->ioc_count;
1456 	} else {
1457 		ASSERT(bp->b_datap->db_type == M_COPYIN);
1458 		count = ((struct copyreq *)bp->b_rptr)->cq_size;
1459 	}
1460 	/*
1461 	 * strdoioctl validates ioc_count, so if this assert fails it
1462 	 * cannot be due to user error.
1463 	 */
1464 	ASSERT(count >= 0);
1465 
1466 	if ((tmp = allocb_cred_wait(count, (flag & STR_NOSIG), &error, cr,
1467 	    curproc->p_pid)) == NULL) {
1468 		return (error);
1469 	}
1470 	error = strcopyin(arg, tmp->b_wptr, count, flag & (U_TO_K|K_TO_K));
1471 	if (error != 0) {
1472 		freeb(tmp);
1473 		return (error);
1474 	}
1475 	DB_CPID(tmp) = curproc->p_pid;
1476 	tmp->b_wptr += count;
1477 	bp->b_cont = tmp;
1478 
1479 	return (0);
1480 }
1481 
1482 /*
1483  * Copy ioctl data to user-land. Return non-zero errno on failure,
1484  * 0 for success.
1485  */
1486 int
getiocd(mblk_t * bp,char * arg,int copymode)1487 getiocd(mblk_t *bp, char *arg, int copymode)
1488 {
1489 	ssize_t count;
1490 	size_t  n;
1491 	int	error;
1492 
1493 	if (bp->b_datap->db_type == M_IOCACK)
1494 		count = ((struct iocblk *)bp->b_rptr)->ioc_count;
1495 	else {
1496 		ASSERT(bp->b_datap->db_type == M_COPYOUT);
1497 		count = ((struct copyreq *)bp->b_rptr)->cq_size;
1498 	}
1499 	ASSERT(count >= 0);
1500 
1501 	for (bp = bp->b_cont; bp && count;
1502 	    count -= n, bp = bp->b_cont, arg += n) {
1503 		n = MIN(count, bp->b_wptr - bp->b_rptr);
1504 		error = strcopyout(bp->b_rptr, arg, n, copymode);
1505 		if (error)
1506 			return (error);
1507 	}
1508 	ASSERT(count == 0);
1509 	return (0);
1510 }
1511 
1512 /*
1513  * Allocate a linkinfo entry given the write queue of the
1514  * bottom module of the top stream and the write queue of the
1515  * stream head of the bottom stream.
1516  */
1517 linkinfo_t *
alloclink(queue_t * qup,queue_t * qdown,file_t * fpdown)1518 alloclink(queue_t *qup, queue_t *qdown, file_t *fpdown)
1519 {
1520 	linkinfo_t *linkp;
1521 
1522 	linkp = kmem_cache_alloc(linkinfo_cache, KM_SLEEP);
1523 
1524 	linkp->li_lblk.l_qtop = qup;
1525 	linkp->li_lblk.l_qbot = qdown;
1526 	linkp->li_fpdown = fpdown;
1527 
1528 	mutex_enter(&strresources);
1529 	linkp->li_next = linkinfo_list;
1530 	linkp->li_prev = NULL;
1531 	if (linkp->li_next)
1532 		linkp->li_next->li_prev = linkp;
1533 	linkinfo_list = linkp;
1534 	linkp->li_lblk.l_index = ++lnk_id;
1535 	ASSERT(lnk_id != 0);	/* this should never wrap in practice */
1536 	mutex_exit(&strresources);
1537 
1538 	return (linkp);
1539 }
1540 
1541 /*
1542  * Free a linkinfo entry.
1543  */
1544 void
lbfree(linkinfo_t * linkp)1545 lbfree(linkinfo_t *linkp)
1546 {
1547 	mutex_enter(&strresources);
1548 	if (linkp->li_next)
1549 		linkp->li_next->li_prev = linkp->li_prev;
1550 	if (linkp->li_prev)
1551 		linkp->li_prev->li_next = linkp->li_next;
1552 	else
1553 		linkinfo_list = linkp->li_next;
1554 	mutex_exit(&strresources);
1555 
1556 	kmem_cache_free(linkinfo_cache, linkp);
1557 }
1558 
1559 /*
1560  * Check for a potential linking cycle.
1561  * Return 1 if a link will result in a cycle,
1562  * and 0 otherwise.
1563  */
1564 int
linkcycle(stdata_t * upstp,stdata_t * lostp,str_stack_t * ss)1565 linkcycle(stdata_t *upstp, stdata_t *lostp, str_stack_t *ss)
1566 {
1567 	struct mux_node *np;
1568 	struct mux_edge *ep;
1569 	int i;
1570 	major_t lomaj;
1571 	major_t upmaj;
1572 	/*
1573 	 * if the lower stream is a pipe/FIFO, return, since link
1574 	 * cycles can not happen on pipes/FIFOs
1575 	 */
1576 	if (lostp->sd_vnode->v_type == VFIFO)
1577 		return (0);
1578 
1579 	for (i = 0; i < ss->ss_devcnt; i++) {
1580 		np = &ss->ss_mux_nodes[i];
1581 		MUX_CLEAR(np);
1582 	}
1583 	lomaj = getmajor(lostp->sd_vnode->v_rdev);
1584 	upmaj = getmajor(upstp->sd_vnode->v_rdev);
1585 	np = &ss->ss_mux_nodes[lomaj];
1586 	for (;;) {
1587 		if (!MUX_DIDVISIT(np)) {
1588 			if (np->mn_imaj == upmaj)
1589 				return (1);
1590 			if (np->mn_outp == NULL) {
1591 				MUX_VISIT(np);
1592 				if (np->mn_originp == NULL)
1593 					return (0);
1594 				np = np->mn_originp;
1595 				continue;
1596 			}
1597 			MUX_VISIT(np);
1598 			np->mn_startp = np->mn_outp;
1599 		} else {
1600 			if (np->mn_startp == NULL) {
1601 				if (np->mn_originp == NULL)
1602 					return (0);
1603 				else {
1604 					np = np->mn_originp;
1605 					continue;
1606 				}
1607 			}
1608 			/*
1609 			 * If ep->me_nodep is a FIFO (me_nodep == NULL),
1610 			 * ignore the edge and move on. ep->me_nodep gets
1611 			 * set to NULL in mux_addedge() if it is a FIFO.
1612 			 *
1613 			 */
1614 			ep = np->mn_startp;
1615 			np->mn_startp = ep->me_nextp;
1616 			if (ep->me_nodep == NULL)
1617 				continue;
1618 			ep->me_nodep->mn_originp = np;
1619 			np = ep->me_nodep;
1620 		}
1621 	}
1622 }
1623 
1624 /*
1625  * Find linkinfo entry corresponding to the parameters.
1626  */
1627 linkinfo_t *
findlinks(stdata_t * stp,int index,int type,str_stack_t * ss)1628 findlinks(stdata_t *stp, int index, int type, str_stack_t *ss)
1629 {
1630 	linkinfo_t *linkp;
1631 	struct mux_edge *mep;
1632 	struct mux_node *mnp;
1633 	queue_t *qup;
1634 
1635 	mutex_enter(&strresources);
1636 	if ((type & LINKTYPEMASK) == LINKNORMAL) {
1637 		qup = getendq(stp->sd_wrq);
1638 		for (linkp = linkinfo_list; linkp; linkp = linkp->li_next) {
1639 			if ((qup == linkp->li_lblk.l_qtop) &&
1640 			    (!index || (index == linkp->li_lblk.l_index))) {
1641 				mutex_exit(&strresources);
1642 				return (linkp);
1643 			}
1644 		}
1645 	} else {
1646 		ASSERT((type & LINKTYPEMASK) == LINKPERSIST);
1647 		mnp = &ss->ss_mux_nodes[getmajor(stp->sd_vnode->v_rdev)];
1648 		mep = mnp->mn_outp;
1649 		while (mep) {
1650 			if ((index == 0) || (index == mep->me_muxid))
1651 				break;
1652 			mep = mep->me_nextp;
1653 		}
1654 		if (!mep) {
1655 			mutex_exit(&strresources);
1656 			return (NULL);
1657 		}
1658 		for (linkp = linkinfo_list; linkp; linkp = linkp->li_next) {
1659 			if ((!linkp->li_lblk.l_qtop) &&
1660 			    (mep->me_muxid == linkp->li_lblk.l_index)) {
1661 				mutex_exit(&strresources);
1662 				return (linkp);
1663 			}
1664 		}
1665 	}
1666 	mutex_exit(&strresources);
1667 	return (NULL);
1668 }
1669 
1670 /*
1671  * Given a queue ptr, follow the chain of q_next pointers until you reach the
1672  * last queue on the chain and return it.
1673  */
1674 queue_t *
getendq(queue_t * q)1675 getendq(queue_t *q)
1676 {
1677 	ASSERT(q != NULL);
1678 	while (_SAMESTR(q))
1679 		q = q->q_next;
1680 	return (q);
1681 }
1682 
1683 /*
1684  * Wait for the syncq count to drop to zero.
1685  * sq could be either outer or inner.
1686  */
1687 
1688 static void
wait_syncq(syncq_t * sq)1689 wait_syncq(syncq_t *sq)
1690 {
1691 	uint16_t count;
1692 
1693 	mutex_enter(SQLOCK(sq));
1694 	count = sq->sq_count;
1695 	SQ_PUTLOCKS_ENTER(sq);
1696 	SUM_SQ_PUTCOUNTS(sq, count);
1697 	while (count != 0) {
1698 		sq->sq_flags |= SQ_WANTWAKEUP;
1699 		SQ_PUTLOCKS_EXIT(sq);
1700 		cv_wait(&sq->sq_wait, SQLOCK(sq));
1701 		count = sq->sq_count;
1702 		SQ_PUTLOCKS_ENTER(sq);
1703 		SUM_SQ_PUTCOUNTS(sq, count);
1704 	}
1705 	SQ_PUTLOCKS_EXIT(sq);
1706 	mutex_exit(SQLOCK(sq));
1707 }
1708 
1709 /*
1710  * Wait while there are any messages for the queue in its syncq.
1711  */
1712 static void
wait_q_syncq(queue_t * q)1713 wait_q_syncq(queue_t *q)
1714 {
1715 	if ((q->q_sqflags & Q_SQQUEUED) || (q->q_syncqmsgs > 0)) {
1716 		syncq_t *sq = q->q_syncq;
1717 
1718 		mutex_enter(SQLOCK(sq));
1719 		while ((q->q_sqflags & Q_SQQUEUED) || (q->q_syncqmsgs > 0)) {
1720 			sq->sq_flags |= SQ_WANTWAKEUP;
1721 			cv_wait(&sq->sq_wait, SQLOCK(sq));
1722 		}
1723 		mutex_exit(SQLOCK(sq));
1724 	}
1725 }
1726 
1727 
1728 int
mlink_file(vnode_t * vp,int cmd,struct file * fpdown,cred_t * crp,int * rvalp,int lhlink)1729 mlink_file(vnode_t *vp, int cmd, struct file *fpdown, cred_t *crp, int *rvalp,
1730     int lhlink)
1731 {
1732 	struct stdata *stp;
1733 	struct strioctl strioc;
1734 	struct linkinfo *linkp;
1735 	struct stdata *stpdown;
1736 	struct streamtab *str;
1737 	queue_t *passq;
1738 	syncq_t *passyncq;
1739 	queue_t *rq;
1740 	cdevsw_impl_t *dp;
1741 	uint32_t qflag;
1742 	uint32_t sqtype;
1743 	perdm_t *dmp;
1744 	int error = 0;
1745 	netstack_t *ns;
1746 	str_stack_t *ss;
1747 
1748 	stp = vp->v_stream;
1749 	TRACE_1(TR_FAC_STREAMS_FR,
1750 	    TR_I_LINK, "I_LINK/I_PLINK:stp %p", stp);
1751 	/*
1752 	 * Test for invalid upper stream
1753 	 */
1754 	if (stp->sd_flag & STRHUP) {
1755 		return (ENXIO);
1756 	}
1757 	if (vp->v_type == VFIFO) {
1758 		return (EINVAL);
1759 	}
1760 	if (stp->sd_strtab == NULL) {
1761 		return (EINVAL);
1762 	}
1763 	if (!stp->sd_strtab->st_muxwinit) {
1764 		return (EINVAL);
1765 	}
1766 	if (fpdown == NULL) {
1767 		return (EBADF);
1768 	}
1769 	ns = netstack_find_by_cred(crp);
1770 	ASSERT(ns != NULL);
1771 	ss = ns->netstack_str;
1772 	ASSERT(ss != NULL);
1773 
1774 	if (getmajor(stp->sd_vnode->v_rdev) >= ss->ss_devcnt) {
1775 		netstack_rele(ss->ss_netstack);
1776 		return (EINVAL);
1777 	}
1778 	mutex_enter(&muxifier);
1779 	if (stp->sd_flag & STPLEX) {
1780 		mutex_exit(&muxifier);
1781 		netstack_rele(ss->ss_netstack);
1782 		return (ENXIO);
1783 	}
1784 
1785 	/*
1786 	 * Test for invalid lower stream.
1787 	 * The check for the v_type != VFIFO and having a major
1788 	 * number not >= devcnt is done to avoid problems with
1789 	 * adding mux_node entry past the end of mux_nodes[].
1790 	 * For FIFO's we don't add an entry so this isn't a
1791 	 * problem.
1792 	 */
1793 	if (((stpdown = fpdown->f_vnode->v_stream) == NULL) ||
1794 	    (stpdown == stp) || (stpdown->sd_flag &
1795 	    (STPLEX|STRHUP|STRDERR|STWRERR|IOCWAIT|STRPLUMB)) ||
1796 	    ((stpdown->sd_vnode->v_type != VFIFO) &&
1797 	    (getmajor(stpdown->sd_vnode->v_rdev) >= ss->ss_devcnt)) ||
1798 	    linkcycle(stp, stpdown, ss)) {
1799 		mutex_exit(&muxifier);
1800 		netstack_rele(ss->ss_netstack);
1801 		return (EINVAL);
1802 	}
1803 	TRACE_1(TR_FAC_STREAMS_FR,
1804 	    TR_STPDOWN, "stpdown:%p", stpdown);
1805 	rq = getendq(stp->sd_wrq);
1806 	if (cmd == I_PLINK)
1807 		rq = NULL;
1808 
1809 	linkp = alloclink(rq, stpdown->sd_wrq, fpdown);
1810 
1811 	strioc.ic_cmd = cmd;
1812 	strioc.ic_timout = INFTIM;
1813 	strioc.ic_len = sizeof (struct linkblk);
1814 	strioc.ic_dp = (char *)&linkp->li_lblk;
1815 
1816 	/*
1817 	 * STRPLUMB protects plumbing changes and should be set before
1818 	 * link_addpassthru()/link_rempassthru() are called, so it is set here
1819 	 * and cleared in the end of mlink when passthru queue is removed.
1820 	 * Setting of STRPLUMB prevents reopens of the stream while passthru
1821 	 * queue is in-place (it is not a proper module and doesn't have open
1822 	 * entry point).
1823 	 *
1824 	 * STPLEX prevents any threads from entering the stream from above. It
1825 	 * can't be set before the call to link_addpassthru() because putnext
1826 	 * from below may cause stream head I/O routines to be called and these
1827 	 * routines assert that STPLEX is not set. After link_addpassthru()
1828 	 * nothing may come from below since the pass queue syncq is blocked.
1829 	 * Note also that STPLEX should be cleared before the call to
1830 	 * link_rempassthru() since when messages start flowing to the stream
1831 	 * head (e.g. because of message propagation from the pass queue) stream
1832 	 * head I/O routines may be called with STPLEX flag set.
1833 	 *
1834 	 * When STPLEX is set, nothing may come into the stream from above and
1835 	 * it is safe to do a setq which will change stream head. So, the
1836 	 * correct sequence of actions is:
1837 	 *
1838 	 * 1) Set STRPLUMB
1839 	 * 2) Call link_addpassthru()
1840 	 * 3) Set STPLEX
1841 	 * 4) Call setq and update the stream state
1842 	 * 5) Clear STPLEX
1843 	 * 6) Call link_rempassthru()
1844 	 * 7) Clear STRPLUMB
1845 	 *
1846 	 * The same sequence applies to munlink() code.
1847 	 */
1848 	mutex_enter(&stpdown->sd_lock);
1849 	stpdown->sd_flag |= STRPLUMB;
1850 	mutex_exit(&stpdown->sd_lock);
1851 	/*
1852 	 * Add passthru queue below lower mux. This will block
1853 	 * syncqs of lower muxs read queue during I_LINK/I_UNLINK.
1854 	 */
1855 	passq = link_addpassthru(stpdown);
1856 
1857 	mutex_enter(&stpdown->sd_lock);
1858 	stpdown->sd_flag |= STPLEX;
1859 	mutex_exit(&stpdown->sd_lock);
1860 
1861 	rq = _RD(stpdown->sd_wrq);
1862 	/*
1863 	 * There may be messages in the streamhead's syncq due to messages
1864 	 * that arrived before link_addpassthru() was done. To avoid
1865 	 * background processing of the syncq happening simultaneous with
1866 	 * setq processing, we disable the streamhead syncq and wait until
1867 	 * existing background thread finishes working on it.
1868 	 */
1869 	wait_sq_svc(rq->q_syncq);
1870 	passyncq = passq->q_syncq;
1871 	if (!(passyncq->sq_flags & SQ_BLOCKED))
1872 		blocksq(passyncq, SQ_BLOCKED, 0);
1873 
1874 	ASSERT((rq->q_flag & QMT_TYPEMASK) == QMTSAFE);
1875 	ASSERT(rq->q_syncq == SQ(rq) && _WR(rq)->q_syncq == SQ(rq));
1876 	rq->q_ptr = _WR(rq)->q_ptr = NULL;
1877 
1878 	/* setq might sleep in allocator - avoid holding locks. */
1879 	/* Note: we are holding muxifier here. */
1880 
1881 	str = stp->sd_strtab;
1882 	dp = &devimpl[getmajor(vp->v_rdev)];
1883 	ASSERT(dp->d_str == str);
1884 
1885 	qflag = dp->d_qflag;
1886 	sqtype = dp->d_sqtype;
1887 
1888 	/* create perdm_t if needed */
1889 	if (NEED_DM(dp->d_dmp, qflag))
1890 		dp->d_dmp = hold_dm(str, qflag, sqtype);
1891 
1892 	dmp = dp->d_dmp;
1893 
1894 	setq(rq, str->st_muxrinit, str->st_muxwinit, dmp, qflag, sqtype,
1895 	    B_TRUE);
1896 
1897 	/*
1898 	 * XXX Remove any "odd" messages from the queue.
1899 	 * Keep only M_DATA, M_PROTO, M_PCPROTO.
1900 	 */
1901 	error = strdoioctl(stp, &strioc, FNATIVE,
1902 	    K_TO_K | STR_NOERROR | STR_NOSIG, crp, rvalp);
1903 	if (error != 0)
1904 		goto cleanup;
1905 
1906 	mutex_enter(&fpdown->f_tlock);
1907 	fpdown->f_count++;
1908 	mutex_exit(&fpdown->f_tlock);
1909 
1910 	/*
1911 	 * if we've made it here the linkage is all set up so we should also
1912 	 * set up the layered driver linkages
1913 	 */
1914 
1915 	ASSERT((cmd == I_LINK) || (cmd == I_PLINK));
1916 	if (cmd == I_LINK) {
1917 		error = ldi_mlink_fp(stp, fpdown, lhlink, LINKNORMAL);
1918 	} else {
1919 		error = ldi_mlink_fp(stp, fpdown, lhlink, LINKPERSIST);
1920 	}
1921 
1922 	if (error != 0) {
1923 		mutex_enter(&fpdown->f_tlock);
1924 		fpdown->f_count--;
1925 		mutex_exit(&fpdown->f_tlock);
1926 		goto cleanup;
1927 	}
1928 
1929 	link_rempassthru(passq);
1930 
1931 	mux_addedge(stp, stpdown, linkp->li_lblk.l_index, ss);
1932 
1933 	/*
1934 	 * Mark the upper stream as having dependent links
1935 	 * so that strclose can clean it up.
1936 	 */
1937 	if (cmd == I_LINK) {
1938 		mutex_enter(&stp->sd_lock);
1939 		stp->sd_flag |= STRHASLINKS;
1940 		mutex_exit(&stp->sd_lock);
1941 	}
1942 	/*
1943 	 * Wake up any other processes that may have been
1944 	 * waiting on the lower stream. These will all
1945 	 * error out.
1946 	 */
1947 	mutex_enter(&stpdown->sd_lock);
1948 	/* The passthru module is removed so we may release STRPLUMB */
1949 	stpdown->sd_flag &= ~STRPLUMB;
1950 	cv_broadcast(&rq->q_wait);
1951 	cv_broadcast(&_WR(rq)->q_wait);
1952 	cv_broadcast(&stpdown->sd_monitor);
1953 	mutex_exit(&stpdown->sd_lock);
1954 	mutex_exit(&muxifier);
1955 	*rvalp = linkp->li_lblk.l_index;
1956 	netstack_rele(ss->ss_netstack);
1957 	return (0);
1958 
1959 cleanup:
1960 	lbfree(linkp);
1961 
1962 	if (!(passyncq->sq_flags & SQ_BLOCKED))
1963 		blocksq(passyncq, SQ_BLOCKED, 0);
1964 	/*
1965 	 * Restore the stream head queue and then remove
1966 	 * the passq. Turn off STPLEX before we turn on
1967 	 * the stream by removing the passq.
1968 	 */
1969 	rq->q_ptr = _WR(rq)->q_ptr = stpdown;
1970 	setq(rq, &strdata, &stwdata, NULL, QMTSAFE, SQ_CI|SQ_CO,
1971 	    B_TRUE);
1972 
1973 	mutex_enter(&stpdown->sd_lock);
1974 	stpdown->sd_flag &= ~STPLEX;
1975 	mutex_exit(&stpdown->sd_lock);
1976 
1977 	link_rempassthru(passq);
1978 
1979 	mutex_enter(&stpdown->sd_lock);
1980 	stpdown->sd_flag &= ~STRPLUMB;
1981 	/* Wakeup anyone waiting for STRPLUMB to clear. */
1982 	cv_broadcast(&stpdown->sd_monitor);
1983 	mutex_exit(&stpdown->sd_lock);
1984 
1985 	mutex_exit(&muxifier);
1986 	netstack_rele(ss->ss_netstack);
1987 	return (error);
1988 }
1989 
1990 int
mlink(vnode_t * vp,int cmd,int arg,cred_t * crp,int * rvalp,int lhlink)1991 mlink(vnode_t *vp, int cmd, int arg, cred_t *crp, int *rvalp, int lhlink)
1992 {
1993 	int		ret;
1994 	struct file	*fpdown;
1995 
1996 	fpdown = getf(arg);
1997 	ret = mlink_file(vp, cmd, fpdown, crp, rvalp, lhlink);
1998 	if (fpdown != NULL)
1999 		releasef(arg);
2000 	return (ret);
2001 }
2002 
2003 /*
2004  * Unlink a multiplexor link. Stp is the controlling stream for the
2005  * link, and linkp points to the link's entry in the linkinfo list.
2006  * The muxifier lock must be held on entry and is dropped on exit.
2007  *
2008  * NOTE : Currently it is assumed that mux would process all the messages
2009  * sitting on it's queue before ACKing the UNLINK. It is the responsibility
2010  * of the mux to handle all the messages that arrive before UNLINK.
2011  * If the mux has to send down messages on its lower stream before
2012  * ACKing I_UNLINK, then it *should* know to handle messages even
2013  * after the UNLINK is acked (actually it should be able to handle till we
2014  * re-block the read side of the pass queue here). If the mux does not
2015  * open up the lower stream, any messages that arrive during UNLINK
2016  * will be put in the stream head. In the case of lower stream opening
2017  * up, some messages might land in the stream head depending on when
2018  * the message arrived and when the read side of the pass queue was
2019  * re-blocked.
2020  */
2021 int
munlink(stdata_t * stp,linkinfo_t * linkp,int flag,cred_t * crp,int * rvalp,str_stack_t * ss)2022 munlink(stdata_t *stp, linkinfo_t *linkp, int flag, cred_t *crp, int *rvalp,
2023     str_stack_t *ss)
2024 {
2025 	struct strioctl strioc;
2026 	struct stdata *stpdown;
2027 	queue_t *rq, *wrq;
2028 	queue_t	*passq;
2029 	syncq_t *passyncq;
2030 	int error = 0;
2031 	file_t *fpdown;
2032 
2033 	ASSERT(MUTEX_HELD(&muxifier));
2034 
2035 	stpdown = linkp->li_fpdown->f_vnode->v_stream;
2036 
2037 	/*
2038 	 * See the comment in mlink() concerning STRPLUMB/STPLEX flags.
2039 	 */
2040 	mutex_enter(&stpdown->sd_lock);
2041 	stpdown->sd_flag |= STRPLUMB;
2042 	mutex_exit(&stpdown->sd_lock);
2043 
2044 	/*
2045 	 * Add passthru queue below lower mux. This will block
2046 	 * syncqs of lower muxs read queue during I_LINK/I_UNLINK.
2047 	 */
2048 	passq = link_addpassthru(stpdown);
2049 
2050 	if ((flag & LINKTYPEMASK) == LINKNORMAL)
2051 		strioc.ic_cmd = I_UNLINK;
2052 	else
2053 		strioc.ic_cmd = I_PUNLINK;
2054 	strioc.ic_timout = INFTIM;
2055 	strioc.ic_len = sizeof (struct linkblk);
2056 	strioc.ic_dp = (char *)&linkp->li_lblk;
2057 
2058 	error = strdoioctl(stp, &strioc, FNATIVE,
2059 	    K_TO_K | STR_NOERROR | STR_NOSIG, crp, rvalp);
2060 
2061 	/*
2062 	 * If there was an error and this is not called via strclose,
2063 	 * return to the user. Otherwise, pretend there was no error
2064 	 * and close the link.
2065 	 */
2066 	if (error) {
2067 		if (flag & LINKCLOSE) {
2068 			cmn_err(CE_WARN, "KERNEL: munlink: could not perform "
2069 			    "unlink ioctl, closing anyway (%d)\n", error);
2070 		} else {
2071 			link_rempassthru(passq);
2072 			mutex_enter(&stpdown->sd_lock);
2073 			stpdown->sd_flag &= ~STRPLUMB;
2074 			cv_broadcast(&stpdown->sd_monitor);
2075 			mutex_exit(&stpdown->sd_lock);
2076 			mutex_exit(&muxifier);
2077 			return (error);
2078 		}
2079 	}
2080 
2081 	mux_rmvedge(stp, linkp->li_lblk.l_index, ss);
2082 	fpdown = linkp->li_fpdown;
2083 	lbfree(linkp);
2084 
2085 	/*
2086 	 * We go ahead and drop muxifier here--it's a nasty global lock that
2087 	 * can slow others down. It's okay to since attempts to mlink() this
2088 	 * stream will be stopped because STPLEX is still set in the stdata
2089 	 * structure, and munlink() is stopped because mux_rmvedge() and
2090 	 * lbfree() have removed it from mux_nodes[] and linkinfo_list,
2091 	 * respectively.  Note that we defer the closef() of fpdown until
2092 	 * after we drop muxifier since strclose() can call munlinkall().
2093 	 */
2094 	mutex_exit(&muxifier);
2095 
2096 	wrq = stpdown->sd_wrq;
2097 	rq = _RD(wrq);
2098 
2099 	/*
2100 	 * Get rid of outstanding service procedure runs, before we make
2101 	 * it a stream head, since a stream head doesn't have any service
2102 	 * procedure.
2103 	 */
2104 	disable_svc(rq);
2105 	wait_svc(rq);
2106 
2107 	/*
2108 	 * Since we don't disable the syncq for QPERMOD, we wait for whatever
2109 	 * is queued up to be finished. mux should take care that nothing is
2110 	 * send down to this queue. We should do it now as we're going to block
2111 	 * passyncq if it was unblocked.
2112 	 */
2113 	if (wrq->q_flag & QPERMOD) {
2114 		syncq_t	*sq = wrq->q_syncq;
2115 
2116 		mutex_enter(SQLOCK(sq));
2117 		while (wrq->q_sqflags & Q_SQQUEUED) {
2118 			sq->sq_flags |= SQ_WANTWAKEUP;
2119 			cv_wait(&sq->sq_wait, SQLOCK(sq));
2120 		}
2121 		mutex_exit(SQLOCK(sq));
2122 	}
2123 	passyncq = passq->q_syncq;
2124 	if (!(passyncq->sq_flags & SQ_BLOCKED)) {
2125 
2126 		syncq_t *sq, *outer;
2127 
2128 		/*
2129 		 * Messages could be flowing from underneath. We will
2130 		 * block the read side of the passq. This would be
2131 		 * sufficient for QPAIR and QPERQ muxes to ensure
2132 		 * that no data is flowing up into this queue
2133 		 * and hence no thread active in this instance of
2134 		 * lower mux. But for QPERMOD and QMTOUTPERIM there
2135 		 * could be messages on the inner and outer/inner
2136 		 * syncqs respectively. We will wait for them to drain.
2137 		 * Because passq is blocked messages end up in the syncq
2138 		 * And qfill_syncq could possibly end up setting QFULL
2139 		 * which will access the rq->q_flag. Hence, we have to
2140 		 * acquire the QLOCK in setq.
2141 		 *
2142 		 * XXX Messages can also flow from top into this
2143 		 * queue though the unlink is over (Ex. some instance
2144 		 * in putnext() called from top that has still not
2145 		 * accessed this queue. And also putq(lowerq) ?).
2146 		 * Solution : How about blocking the l_qtop queue ?
2147 		 * Do we really care about such pure D_MP muxes ?
2148 		 */
2149 
2150 		blocksq(passyncq, SQ_BLOCKED, 0);
2151 
2152 		sq = rq->q_syncq;
2153 		if ((outer = sq->sq_outer) != NULL) {
2154 
2155 			/*
2156 			 * We have to just wait for the outer sq_count
2157 			 * drop to zero. As this does not prevent new
2158 			 * messages to enter the outer perimeter, this
2159 			 * is subject to starvation.
2160 			 *
2161 			 * NOTE :Because of blocksq above, messages could
2162 			 * be in the inner syncq only because of some
2163 			 * thread holding the outer perimeter exclusively.
2164 			 * Hence it would be sufficient to wait for the
2165 			 * exclusive holder of the outer perimeter to drain
2166 			 * the inner and outer syncqs. But we will not depend
2167 			 * on this feature and hence check the inner syncqs
2168 			 * separately.
2169 			 */
2170 			wait_syncq(outer);
2171 		}
2172 
2173 
2174 		/*
2175 		 * There could be messages destined for
2176 		 * this queue. Let the exclusive holder
2177 		 * drain it.
2178 		 */
2179 
2180 		wait_syncq(sq);
2181 		ASSERT((rq->q_flag & QPERMOD) ||
2182 		    ((rq->q_syncq->sq_head == NULL) &&
2183 		    (_WR(rq)->q_syncq->sq_head == NULL)));
2184 	}
2185 
2186 	/*
2187 	 * We haven't taken care of QPERMOD case yet. QPERMOD is a special
2188 	 * case as we don't disable its syncq or remove it off the syncq
2189 	 * service list.
2190 	 */
2191 	if (rq->q_flag & QPERMOD) {
2192 		syncq_t	*sq = rq->q_syncq;
2193 
2194 		mutex_enter(SQLOCK(sq));
2195 		while (rq->q_sqflags & Q_SQQUEUED) {
2196 			sq->sq_flags |= SQ_WANTWAKEUP;
2197 			cv_wait(&sq->sq_wait, SQLOCK(sq));
2198 		}
2199 		mutex_exit(SQLOCK(sq));
2200 	}
2201 
2202 	/*
2203 	 * flush_syncq changes states only when there are some messages to
2204 	 * free, i.e. when it returns non-zero value to return.
2205 	 */
2206 	ASSERT(flush_syncq(rq->q_syncq, rq) == 0);
2207 	ASSERT(flush_syncq(wrq->q_syncq, wrq) == 0);
2208 
2209 	/*
2210 	 * Nobody else should know about this queue now.
2211 	 * If the mux did not process the messages before
2212 	 * acking the I_UNLINK, free them now.
2213 	 */
2214 
2215 	flushq(rq, FLUSHALL);
2216 	flushq(_WR(rq), FLUSHALL);
2217 
2218 	/*
2219 	 * Convert the mux lower queue into a stream head queue.
2220 	 * Turn off STPLEX before we turn on the stream by removing the passq.
2221 	 */
2222 	rq->q_ptr = wrq->q_ptr = stpdown;
2223 	setq(rq, &strdata, &stwdata, NULL, QMTSAFE, SQ_CI|SQ_CO, B_TRUE);
2224 
2225 	ASSERT((rq->q_flag & QMT_TYPEMASK) == QMTSAFE);
2226 	ASSERT(rq->q_syncq == SQ(rq) && _WR(rq)->q_syncq == SQ(rq));
2227 
2228 	enable_svc(rq);
2229 
2230 	/*
2231 	 * Now it is a proper stream, so STPLEX is cleared. But STRPLUMB still
2232 	 * needs to be set to prevent reopen() of the stream - such reopen may
2233 	 * try to call non-existent pass queue open routine and panic.
2234 	 */
2235 	mutex_enter(&stpdown->sd_lock);
2236 	stpdown->sd_flag &= ~STPLEX;
2237 	mutex_exit(&stpdown->sd_lock);
2238 
2239 	ASSERT(((flag & LINKTYPEMASK) == LINKNORMAL) ||
2240 	    ((flag & LINKTYPEMASK) == LINKPERSIST));
2241 
2242 	/* clean up the layered driver linkages */
2243 	if ((flag & LINKTYPEMASK) == LINKNORMAL) {
2244 		VERIFY0(ldi_munlink_fp(stp, fpdown, LINKNORMAL));
2245 	} else {
2246 		VERIFY0(ldi_munlink_fp(stp, fpdown, LINKPERSIST));
2247 	}
2248 
2249 	link_rempassthru(passq);
2250 
2251 	/*
2252 	 * Now all plumbing changes are finished and STRPLUMB is no
2253 	 * longer needed.
2254 	 */
2255 	mutex_enter(&stpdown->sd_lock);
2256 	stpdown->sd_flag &= ~STRPLUMB;
2257 	cv_broadcast(&stpdown->sd_monitor);
2258 	mutex_exit(&stpdown->sd_lock);
2259 
2260 	(void) closef(fpdown);
2261 	return (0);
2262 }
2263 
2264 /*
2265  * Unlink all multiplexor links for which stp is the controlling stream.
2266  * Return 0, or a non-zero errno on failure.
2267  */
2268 int
munlinkall(stdata_t * stp,int flag,cred_t * crp,int * rvalp,str_stack_t * ss)2269 munlinkall(stdata_t *stp, int flag, cred_t *crp, int *rvalp, str_stack_t *ss)
2270 {
2271 	linkinfo_t *linkp;
2272 	int error = 0;
2273 
2274 	mutex_enter(&muxifier);
2275 	while (linkp = findlinks(stp, 0, flag, ss)) {
2276 		/*
2277 		 * munlink() releases the muxifier lock.
2278 		 */
2279 		if (error = munlink(stp, linkp, flag, crp, rvalp, ss))
2280 			return (error);
2281 		mutex_enter(&muxifier);
2282 	}
2283 	mutex_exit(&muxifier);
2284 	return (0);
2285 }
2286 
2287 /*
2288  * A multiplexor link has been made. Add an
2289  * edge to the directed graph.
2290  */
2291 void
mux_addedge(stdata_t * upstp,stdata_t * lostp,int muxid,str_stack_t * ss)2292 mux_addedge(stdata_t *upstp, stdata_t *lostp, int muxid, str_stack_t *ss)
2293 {
2294 	struct mux_node *np;
2295 	struct mux_edge *ep;
2296 	major_t upmaj;
2297 	major_t lomaj;
2298 
2299 	upmaj = getmajor(upstp->sd_vnode->v_rdev);
2300 	lomaj = getmajor(lostp->sd_vnode->v_rdev);
2301 	np = &ss->ss_mux_nodes[upmaj];
2302 	if (np->mn_outp) {
2303 		ep = np->mn_outp;
2304 		while (ep->me_nextp)
2305 			ep = ep->me_nextp;
2306 		ep->me_nextp = kmem_alloc(sizeof (struct mux_edge), KM_SLEEP);
2307 		ep = ep->me_nextp;
2308 	} else {
2309 		np->mn_outp = kmem_alloc(sizeof (struct mux_edge), KM_SLEEP);
2310 		ep = np->mn_outp;
2311 	}
2312 	ep->me_nextp = NULL;
2313 	ep->me_muxid = muxid;
2314 	/*
2315 	 * Save the dev_t for the purposes of str_stack_shutdown.
2316 	 * str_stack_shutdown assumes that the device allows reopen, since
2317 	 * this dev_t is the one after any cloning by xx_open().
2318 	 * Would prefer finding the dev_t from before any cloning,
2319 	 * but specfs doesn't retain that.
2320 	 */
2321 	ep->me_dev = upstp->sd_vnode->v_rdev;
2322 	if (lostp->sd_vnode->v_type == VFIFO)
2323 		ep->me_nodep = NULL;
2324 	else
2325 		ep->me_nodep = &ss->ss_mux_nodes[lomaj];
2326 }
2327 
2328 /*
2329  * A multiplexor link has been removed. Remove the
2330  * edge in the directed graph.
2331  */
2332 void
mux_rmvedge(stdata_t * upstp,int muxid,str_stack_t * ss)2333 mux_rmvedge(stdata_t *upstp, int muxid, str_stack_t *ss)
2334 {
2335 	struct mux_node *np;
2336 	struct mux_edge *ep;
2337 	struct mux_edge *pep = NULL;
2338 	major_t upmaj;
2339 
2340 	upmaj = getmajor(upstp->sd_vnode->v_rdev);
2341 	np = &ss->ss_mux_nodes[upmaj];
2342 	ASSERT(np->mn_outp != NULL);
2343 	ep = np->mn_outp;
2344 	while (ep) {
2345 		if (ep->me_muxid == muxid) {
2346 			if (pep)
2347 				pep->me_nextp = ep->me_nextp;
2348 			else
2349 				np->mn_outp = ep->me_nextp;
2350 			kmem_free(ep, sizeof (struct mux_edge));
2351 			return;
2352 		}
2353 		pep = ep;
2354 		ep = ep->me_nextp;
2355 	}
2356 	ASSERT(0);	/* should not reach here */
2357 }
2358 
2359 /*
2360  * Translate the device flags (from conf.h) to the corresponding
2361  * qflag and sq_flag (type) values.
2362  */
2363 int
devflg_to_qflag(struct streamtab * stp,uint32_t devflag,uint32_t * qflagp,uint32_t * sqtypep)2364 devflg_to_qflag(struct streamtab *stp, uint32_t devflag, uint32_t *qflagp,
2365     uint32_t *sqtypep)
2366 {
2367 	uint32_t qflag = 0;
2368 	uint32_t sqtype = 0;
2369 
2370 	if (devflag & _D_OLD)
2371 		goto bad;
2372 
2373 	/* Inner perimeter presence and scope */
2374 	switch (devflag & D_MTINNER_MASK) {
2375 	case D_MP:
2376 		qflag |= QMTSAFE;
2377 		sqtype |= SQ_CI;
2378 		break;
2379 	case D_MTPERQ|D_MP:
2380 		qflag |= QPERQ;
2381 		break;
2382 	case D_MTQPAIR|D_MP:
2383 		qflag |= QPAIR;
2384 		break;
2385 	case D_MTPERMOD|D_MP:
2386 		qflag |= QPERMOD;
2387 		break;
2388 	default:
2389 		goto bad;
2390 	}
2391 
2392 	/* Outer perimeter */
2393 	if (devflag & D_MTOUTPERIM) {
2394 		switch (devflag & D_MTINNER_MASK) {
2395 		case D_MP:
2396 		case D_MTPERQ|D_MP:
2397 		case D_MTQPAIR|D_MP:
2398 			break;
2399 		default:
2400 			goto bad;
2401 		}
2402 		qflag |= QMTOUTPERIM;
2403 	}
2404 
2405 	/* Inner perimeter modifiers */
2406 	if (devflag & D_MTINNER_MOD) {
2407 		switch (devflag & D_MTINNER_MASK) {
2408 		case D_MP:
2409 			goto bad;
2410 		default:
2411 			break;
2412 		}
2413 		if (devflag & D_MTPUTSHARED)
2414 			sqtype |= SQ_CIPUT;
2415 		if (devflag & _D_MTOCSHARED) {
2416 			/*
2417 			 * The code in putnext assumes that it has the
2418 			 * highest concurrency by not checking sq_count.
2419 			 * Thus _D_MTOCSHARED can only be supported when
2420 			 * D_MTPUTSHARED is set.
2421 			 */
2422 			if (!(devflag & D_MTPUTSHARED))
2423 				goto bad;
2424 			sqtype |= SQ_CIOC;
2425 		}
2426 		if (devflag & _D_MTCBSHARED) {
2427 			/*
2428 			 * The code in putnext assumes that it has the
2429 			 * highest concurrency by not checking sq_count.
2430 			 * Thus _D_MTCBSHARED can only be supported when
2431 			 * D_MTPUTSHARED is set.
2432 			 */
2433 			if (!(devflag & D_MTPUTSHARED))
2434 				goto bad;
2435 			sqtype |= SQ_CICB;
2436 		}
2437 		if (devflag & _D_MTSVCSHARED) {
2438 			/*
2439 			 * The code in putnext assumes that it has the
2440 			 * highest concurrency by not checking sq_count.
2441 			 * Thus _D_MTSVCSHARED can only be supported when
2442 			 * D_MTPUTSHARED is set. Also _D_MTSVCSHARED is
2443 			 * supported only for QPERMOD.
2444 			 */
2445 			if (!(devflag & D_MTPUTSHARED) || !(qflag & QPERMOD))
2446 				goto bad;
2447 			sqtype |= SQ_CISVC;
2448 		}
2449 	}
2450 
2451 	/* Default outer perimeter concurrency */
2452 	sqtype |= SQ_CO;
2453 
2454 	/* Outer perimeter modifiers */
2455 	if (devflag & D_MTOCEXCL) {
2456 		if (!(devflag & D_MTOUTPERIM)) {
2457 			/* No outer perimeter */
2458 			goto bad;
2459 		}
2460 		sqtype &= ~SQ_COOC;
2461 	}
2462 
2463 	/* Synchronous Streams extended qinit structure */
2464 	if (devflag & D_SYNCSTR)
2465 		qflag |= QSYNCSTR;
2466 
2467 	/*
2468 	 * Private flag used by a transport module to indicate
2469 	 * to sockfs that it supports direct-access mode without
2470 	 * having to go through STREAMS.
2471 	 */
2472 	if (devflag & _D_DIRECT) {
2473 		/* Reject unless the module is fully-MT (no perimeter) */
2474 		if ((qflag & QMT_TYPEMASK) != QMTSAFE)
2475 			goto bad;
2476 		qflag |= _QDIRECT;
2477 	}
2478 
2479 	/*
2480 	 * Private flag used to indicate that a streams module should only
2481 	 * be pushed once. The TTY streams modules have this flag since if
2482 	 * libc believes itself to be an xpg4 process then it will
2483 	 * automatically and unconditionally push them when a PTS device is
2484 	 * opened. If an application is not aware of this then without this
2485 	 * flag we would end up with duplicate modules.
2486 	 */
2487 	if (devflag & _D_SINGLE_INSTANCE)
2488 		qflag |= _QSINGLE_INSTANCE;
2489 
2490 	*qflagp = qflag;
2491 	*sqtypep = sqtype;
2492 	return (0);
2493 
2494 bad:
2495 	cmn_err(CE_WARN,
2496 	    "stropen: bad MT flags (0x%x) in driver '%s'",
2497 	    (int)(qflag & D_MTSAFETY_MASK),
2498 	    stp->st_rdinit->qi_minfo->mi_idname);
2499 
2500 	return (EINVAL);
2501 }
2502 
2503 /*
2504  * Set the interface values for a pair of queues (qinit structure,
2505  * packet sizes, water marks).
2506  * setq assumes that the caller does not have a claim (entersq or claimq)
2507  * on the queue.
2508  */
2509 void
setq(queue_t * rq,struct qinit * rinit,struct qinit * winit,perdm_t * dmp,uint32_t qflag,uint32_t sqtype,boolean_t lock_needed)2510 setq(queue_t *rq, struct qinit *rinit, struct qinit *winit,
2511     perdm_t *dmp, uint32_t qflag, uint32_t sqtype, boolean_t lock_needed)
2512 {
2513 	queue_t *wq;
2514 	syncq_t	*sq, *outer;
2515 
2516 	ASSERT(rq->q_flag & QREADR);
2517 	ASSERT((qflag & QMT_TYPEMASK) != 0);
2518 	IMPLY((qflag & (QPERMOD | QMTOUTPERIM)), dmp != NULL);
2519 
2520 	wq = _WR(rq);
2521 	rq->q_qinfo = rinit;
2522 	rq->q_hiwat = rinit->qi_minfo->mi_hiwat;
2523 	rq->q_lowat = rinit->qi_minfo->mi_lowat;
2524 	rq->q_minpsz = rinit->qi_minfo->mi_minpsz;
2525 	rq->q_maxpsz = rinit->qi_minfo->mi_maxpsz;
2526 	wq->q_qinfo = winit;
2527 	wq->q_hiwat = winit->qi_minfo->mi_hiwat;
2528 	wq->q_lowat = winit->qi_minfo->mi_lowat;
2529 	wq->q_minpsz = winit->qi_minfo->mi_minpsz;
2530 	wq->q_maxpsz = winit->qi_minfo->mi_maxpsz;
2531 
2532 	/* Remove old syncqs */
2533 	sq = rq->q_syncq;
2534 	outer = sq->sq_outer;
2535 	if (outer != NULL) {
2536 		ASSERT(wq->q_syncq->sq_outer == outer);
2537 		outer_remove(outer, rq->q_syncq);
2538 		if (wq->q_syncq != rq->q_syncq)
2539 			outer_remove(outer, wq->q_syncq);
2540 	}
2541 	ASSERT(sq->sq_outer == NULL);
2542 	ASSERT(sq->sq_onext == NULL && sq->sq_oprev == NULL);
2543 
2544 	if (sq != SQ(rq)) {
2545 		if (!(rq->q_flag & QPERMOD))
2546 			free_syncq(sq);
2547 		if (wq->q_syncq == rq->q_syncq)
2548 			wq->q_syncq = NULL;
2549 		rq->q_syncq = NULL;
2550 	}
2551 	if (wq->q_syncq != NULL && wq->q_syncq != sq &&
2552 	    wq->q_syncq != SQ(rq)) {
2553 		free_syncq(wq->q_syncq);
2554 		wq->q_syncq = NULL;
2555 	}
2556 	ASSERT(rq->q_syncq == NULL || (rq->q_syncq->sq_head == NULL &&
2557 	    rq->q_syncq->sq_tail == NULL));
2558 	ASSERT(wq->q_syncq == NULL || (wq->q_syncq->sq_head == NULL &&
2559 	    wq->q_syncq->sq_tail == NULL));
2560 
2561 	if (!(rq->q_flag & QPERMOD) &&
2562 	    rq->q_syncq != NULL && rq->q_syncq->sq_ciputctrl != NULL) {
2563 		ASSERT(rq->q_syncq->sq_nciputctrl == n_ciputctrl - 1);
2564 		SUMCHECK_CIPUTCTRL_COUNTS(rq->q_syncq->sq_ciputctrl,
2565 		    rq->q_syncq->sq_nciputctrl, 0);
2566 		ASSERT(ciputctrl_cache != NULL);
2567 		kmem_cache_free(ciputctrl_cache, rq->q_syncq->sq_ciputctrl);
2568 		rq->q_syncq->sq_ciputctrl = NULL;
2569 		rq->q_syncq->sq_nciputctrl = 0;
2570 	}
2571 
2572 	if (!(wq->q_flag & QPERMOD) &&
2573 	    wq->q_syncq != NULL && wq->q_syncq->sq_ciputctrl != NULL) {
2574 		ASSERT(wq->q_syncq->sq_nciputctrl == n_ciputctrl - 1);
2575 		SUMCHECK_CIPUTCTRL_COUNTS(wq->q_syncq->sq_ciputctrl,
2576 		    wq->q_syncq->sq_nciputctrl, 0);
2577 		ASSERT(ciputctrl_cache != NULL);
2578 		kmem_cache_free(ciputctrl_cache, wq->q_syncq->sq_ciputctrl);
2579 		wq->q_syncq->sq_ciputctrl = NULL;
2580 		wq->q_syncq->sq_nciputctrl = 0;
2581 	}
2582 
2583 	sq = SQ(rq);
2584 	ASSERT(sq->sq_head == NULL && sq->sq_tail == NULL);
2585 	ASSERT(sq->sq_outer == NULL);
2586 	ASSERT(sq->sq_onext == NULL && sq->sq_oprev == NULL);
2587 
2588 	/*
2589 	 * Create syncqs based on qflag and sqtype. Set the SQ_TYPES_IN_FLAGS
2590 	 * bits in sq_flag based on the sqtype.
2591 	 */
2592 	ASSERT((sq->sq_flags & ~SQ_TYPES_IN_FLAGS) == 0);
2593 
2594 	rq->q_syncq = wq->q_syncq = sq;
2595 	sq->sq_type = sqtype;
2596 	sq->sq_flags = (sqtype & SQ_TYPES_IN_FLAGS);
2597 
2598 	/*
2599 	 *  We are making sq_svcflags zero,
2600 	 *  resetting SQ_DISABLED in case it was set by
2601 	 *  wait_svc() in the munlink path.
2602 	 *
2603 	 */
2604 	ASSERT((sq->sq_svcflags & SQ_SERVICE) == 0);
2605 	sq->sq_svcflags = 0;
2606 
2607 	/*
2608 	 * We need to acquire the lock here for the mlink and munlink case,
2609 	 * where canputnext, backenable, etc can access the q_flag.
2610 	 */
2611 	if (lock_needed) {
2612 		mutex_enter(QLOCK(rq));
2613 		rq->q_flag = (rq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2614 		mutex_exit(QLOCK(rq));
2615 		mutex_enter(QLOCK(wq));
2616 		wq->q_flag = (wq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2617 		mutex_exit(QLOCK(wq));
2618 	} else {
2619 		rq->q_flag = (rq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2620 		wq->q_flag = (wq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2621 	}
2622 
2623 	if (qflag & QPERQ) {
2624 		/* Allocate a separate syncq for the write side */
2625 		sq = new_syncq();
2626 		sq->sq_type = rq->q_syncq->sq_type;
2627 		sq->sq_flags = rq->q_syncq->sq_flags;
2628 		ASSERT(sq->sq_outer == NULL && sq->sq_onext == NULL &&
2629 		    sq->sq_oprev == NULL);
2630 		wq->q_syncq = sq;
2631 	}
2632 	if (qflag & QPERMOD) {
2633 		sq = dmp->dm_sq;
2634 
2635 		/*
2636 		 * Assert that we do have an inner perimeter syncq and that it
2637 		 * does not have an outer perimeter associated with it.
2638 		 */
2639 		ASSERT(sq->sq_outer == NULL && sq->sq_onext == NULL &&
2640 		    sq->sq_oprev == NULL);
2641 		rq->q_syncq = wq->q_syncq = sq;
2642 	}
2643 	if (qflag & QMTOUTPERIM) {
2644 		outer = dmp->dm_sq;
2645 
2646 		ASSERT(outer->sq_outer == NULL);
2647 		outer_insert(outer, rq->q_syncq);
2648 		if (wq->q_syncq != rq->q_syncq)
2649 			outer_insert(outer, wq->q_syncq);
2650 	}
2651 	ASSERT((rq->q_syncq->sq_flags & SQ_TYPES_IN_FLAGS) ==
2652 	    (rq->q_syncq->sq_type & SQ_TYPES_IN_FLAGS));
2653 	ASSERT((wq->q_syncq->sq_flags & SQ_TYPES_IN_FLAGS) ==
2654 	    (wq->q_syncq->sq_type & SQ_TYPES_IN_FLAGS));
2655 	ASSERT((rq->q_flag & QMT_TYPEMASK) == (qflag & QMT_TYPEMASK));
2656 
2657 	/*
2658 	 * Initialize struio() types.
2659 	 */
2660 	rq->q_struiot =
2661 	    (rq->q_flag & QSYNCSTR) ? rinit->qi_struiot : STRUIOT_NONE;
2662 	wq->q_struiot =
2663 	    (wq->q_flag & QSYNCSTR) ? winit->qi_struiot : STRUIOT_NONE;
2664 }
2665 
2666 perdm_t *
hold_dm(struct streamtab * str,uint32_t qflag,uint32_t sqtype)2667 hold_dm(struct streamtab *str, uint32_t qflag, uint32_t sqtype)
2668 {
2669 	syncq_t	*sq;
2670 	perdm_t	**pp;
2671 	perdm_t	*p;
2672 	perdm_t	*dmp;
2673 
2674 	ASSERT(str != NULL);
2675 	ASSERT(qflag & (QPERMOD | QMTOUTPERIM));
2676 
2677 	rw_enter(&perdm_rwlock, RW_READER);
2678 	for (p = perdm_list; p != NULL; p = p->dm_next) {
2679 		if (p->dm_str == str) {	/* found one */
2680 			atomic_inc_32(&(p->dm_ref));
2681 			rw_exit(&perdm_rwlock);
2682 			return (p);
2683 		}
2684 	}
2685 	rw_exit(&perdm_rwlock);
2686 
2687 	sq = new_syncq();
2688 	if (qflag & QPERMOD) {
2689 		sq->sq_type = sqtype | SQ_PERMOD;
2690 		sq->sq_flags = sqtype & SQ_TYPES_IN_FLAGS;
2691 	} else {
2692 		ASSERT(qflag & QMTOUTPERIM);
2693 		sq->sq_onext = sq->sq_oprev = sq;
2694 	}
2695 
2696 	dmp = kmem_alloc(sizeof (perdm_t), KM_SLEEP);
2697 	dmp->dm_sq = sq;
2698 	dmp->dm_str = str;
2699 	dmp->dm_ref = 1;
2700 	dmp->dm_next = NULL;
2701 
2702 	rw_enter(&perdm_rwlock, RW_WRITER);
2703 	for (pp = &perdm_list; (p = *pp) != NULL; pp = &(p->dm_next)) {
2704 		if (p->dm_str == str) {	/* already present */
2705 			p->dm_ref++;
2706 			rw_exit(&perdm_rwlock);
2707 			free_syncq(sq);
2708 			kmem_free(dmp, sizeof (perdm_t));
2709 			return (p);
2710 		}
2711 	}
2712 
2713 	*pp = dmp;
2714 	rw_exit(&perdm_rwlock);
2715 	return (dmp);
2716 }
2717 
2718 void
rele_dm(perdm_t * dmp)2719 rele_dm(perdm_t *dmp)
2720 {
2721 	perdm_t **pp;
2722 	perdm_t *p;
2723 
2724 	rw_enter(&perdm_rwlock, RW_WRITER);
2725 	ASSERT(dmp->dm_ref > 0);
2726 
2727 	if (--dmp->dm_ref > 0) {
2728 		rw_exit(&perdm_rwlock);
2729 		return;
2730 	}
2731 
2732 	for (pp = &perdm_list; (p = *pp) != NULL; pp = &(p->dm_next))
2733 		if (p == dmp)
2734 			break;
2735 	ASSERT(p == dmp);
2736 	*pp = p->dm_next;
2737 	rw_exit(&perdm_rwlock);
2738 
2739 	/*
2740 	 * Wait for any background processing that relies on the
2741 	 * syncq to complete before it is freed.
2742 	 */
2743 	wait_sq_svc(p->dm_sq);
2744 	free_syncq(p->dm_sq);
2745 	kmem_free(p, sizeof (perdm_t));
2746 }
2747 
2748 /*
2749  * Make a protocol message given control and data buffers.
2750  * n.b., this can block; be careful of what locks you hold when calling it.
2751  *
2752  * If sd_maxblk is less than *iosize this routine can fail part way through
2753  * (due to an allocation failure). In this case on return *iosize will contain
2754  * the amount that was consumed. Otherwise *iosize will not be modified
2755  * i.e. it will contain the amount that was consumed.
2756  */
2757 int
strmakemsg(struct strbuf * mctl,ssize_t * iosize,struct uio * uiop,stdata_t * stp,int32_t flag,mblk_t ** mpp)2758 strmakemsg(
2759 	struct strbuf *mctl,
2760 	ssize_t *iosize,
2761 	struct uio *uiop,
2762 	stdata_t *stp,
2763 	int32_t flag,
2764 	mblk_t **mpp)
2765 {
2766 	mblk_t *mpctl = NULL;
2767 	mblk_t *mpdata = NULL;
2768 	int error;
2769 
2770 	ASSERT(uiop != NULL);
2771 
2772 	*mpp = NULL;
2773 	/* Create control part, if any */
2774 	if ((mctl != NULL) && (mctl->len >= 0)) {
2775 		error = strmakectl(mctl, flag, uiop->uio_fmode, &mpctl);
2776 		if (error)
2777 			return (error);
2778 	}
2779 	/* Create data part, if any */
2780 	if (*iosize >= 0) {
2781 		error = strmakedata(iosize, uiop, stp, flag, &mpdata);
2782 		if (error) {
2783 			freemsg(mpctl);
2784 			return (error);
2785 		}
2786 	}
2787 	if (mpctl != NULL) {
2788 		if (mpdata != NULL)
2789 			linkb(mpctl, mpdata);
2790 		*mpp = mpctl;
2791 	} else {
2792 		*mpp = mpdata;
2793 	}
2794 	return (0);
2795 }
2796 
2797 /*
2798  * Make the control part of a protocol message given a control buffer.
2799  * n.b., this can block; be careful of what locks you hold when calling it.
2800  */
2801 int
strmakectl(struct strbuf * mctl,int32_t flag,int32_t fflag,mblk_t ** mpp)2802 strmakectl(
2803 	struct strbuf *mctl,
2804 	int32_t flag,
2805 	int32_t fflag,
2806 	mblk_t **mpp)
2807 {
2808 	mblk_t *bp = NULL;
2809 	unsigned char msgtype;
2810 	int error = 0;
2811 	cred_t *cr = CRED();
2812 
2813 	/* We do not support interrupt threads using the stream head to send */
2814 	ASSERT(cr != NULL);
2815 
2816 	*mpp = NULL;
2817 	/*
2818 	 * Create control part of message, if any.
2819 	 */
2820 	if ((mctl != NULL) && (mctl->len >= 0)) {
2821 		caddr_t base;
2822 		int ctlcount;
2823 		int allocsz;
2824 
2825 		if (flag & RS_HIPRI)
2826 			msgtype = M_PCPROTO;
2827 		else
2828 			msgtype = M_PROTO;
2829 
2830 		ctlcount = mctl->len;
2831 		base = mctl->buf;
2832 
2833 		/*
2834 		 * Give modules a better chance to reuse M_PROTO/M_PCPROTO
2835 		 * blocks by increasing the size to something more usable.
2836 		 */
2837 		allocsz = MAX(ctlcount, 64);
2838 
2839 		/*
2840 		 * Range checking has already been done; simply try
2841 		 * to allocate a message block for the ctl part.
2842 		 */
2843 		while ((bp = allocb_cred(allocsz, cr,
2844 		    curproc->p_pid)) == NULL) {
2845 			if (fflag & (FNDELAY|FNONBLOCK))
2846 				return (EAGAIN);
2847 			if (error = strwaitbuf(allocsz, BPRI_MED))
2848 				return (error);
2849 		}
2850 
2851 		bp->b_datap->db_type = msgtype;
2852 		if (copyin(base, bp->b_wptr, ctlcount)) {
2853 			freeb(bp);
2854 			return (EFAULT);
2855 		}
2856 		bp->b_wptr += ctlcount;
2857 	}
2858 	*mpp = bp;
2859 	return (0);
2860 }
2861 
2862 /*
2863  * Make a protocol message given data buffers.
2864  * n.b., this can block; be careful of what locks you hold when calling it.
2865  *
2866  * If sd_maxblk is less than *iosize this routine can fail part way through
2867  * (due to an allocation failure). In this case on return *iosize will contain
2868  * the amount that was consumed. Otherwise *iosize will not be modified
2869  * i.e. it will contain the amount that was consumed.
2870  */
2871 int
strmakedata(ssize_t * iosize,struct uio * uiop,stdata_t * stp,int32_t flag,mblk_t ** mpp)2872 strmakedata(
2873 	ssize_t   *iosize,
2874 	struct uio *uiop,
2875 	stdata_t *stp,
2876 	int32_t flag,
2877 	mblk_t **mpp)
2878 {
2879 	mblk_t *mp = NULL;
2880 	mblk_t *bp;
2881 	int wroff = (int)stp->sd_wroff;
2882 	int tail_len = (int)stp->sd_tail;
2883 	int extra = wroff + tail_len;
2884 	int error = 0;
2885 	ssize_t maxblk;
2886 	ssize_t count = *iosize;
2887 	cred_t *cr;
2888 
2889 	*mpp = NULL;
2890 	if (count < 0)
2891 		return (0);
2892 
2893 	/* We do not support interrupt threads using the stream head to send */
2894 	cr = CRED();
2895 	ASSERT(cr != NULL);
2896 
2897 	maxblk = stp->sd_maxblk;
2898 	if (maxblk == INFPSZ)
2899 		maxblk = count;
2900 
2901 	/*
2902 	 * Create data part of message, if any.
2903 	 */
2904 	do {
2905 		ssize_t size;
2906 		dblk_t  *dp;
2907 
2908 		ASSERT(uiop);
2909 
2910 		size = MIN(count, maxblk);
2911 
2912 		while ((bp = allocb_cred(size + extra, cr,
2913 		    curproc->p_pid)) == NULL) {
2914 			error = EAGAIN;
2915 			if ((uiop->uio_fmode & (FNDELAY|FNONBLOCK)) ||
2916 			    (error = strwaitbuf(size + extra, BPRI_MED)) != 0) {
2917 				if (count == *iosize) {
2918 					freemsg(mp);
2919 					return (error);
2920 				} else {
2921 					*iosize -= count;
2922 					*mpp = mp;
2923 					return (0);
2924 				}
2925 			}
2926 		}
2927 		dp = bp->b_datap;
2928 		dp->db_cpid = curproc->p_pid;
2929 		ASSERT(wroff <= dp->db_lim - bp->b_wptr);
2930 		bp->b_wptr = bp->b_rptr = bp->b_rptr + wroff;
2931 
2932 		if (flag & STRUIO_POSTPONE) {
2933 			/*
2934 			 * Setup the stream uio portion of the
2935 			 * dblk for subsequent use by struioget().
2936 			 */
2937 			dp->db_struioflag = STRUIO_SPEC;
2938 			dp->db_cksumstart = 0;
2939 			dp->db_cksumstuff = 0;
2940 			dp->db_cksumend = size;
2941 			*(long long *)dp->db_struioun.data = 0ll;
2942 			bp->b_wptr += size;
2943 		} else {
2944 			if (stp->sd_copyflag & STRCOPYCACHED)
2945 				uiop->uio_extflg |= UIO_COPY_CACHED;
2946 
2947 			if (size != 0) {
2948 				error = uiomove(bp->b_wptr, size, UIO_WRITE,
2949 				    uiop);
2950 				if (error != 0) {
2951 					freeb(bp);
2952 					freemsg(mp);
2953 					return (error);
2954 				}
2955 			}
2956 			bp->b_wptr += size;
2957 
2958 			if (stp->sd_wputdatafunc != NULL) {
2959 				mblk_t *newbp;
2960 
2961 				newbp = (stp->sd_wputdatafunc)(stp->sd_vnode,
2962 				    bp, NULL, NULL, NULL, NULL);
2963 				if (newbp == NULL) {
2964 					freeb(bp);
2965 					freemsg(mp);
2966 					return (ECOMM);
2967 				}
2968 				bp = newbp;
2969 			}
2970 		}
2971 
2972 		count -= size;
2973 
2974 		if (mp == NULL)
2975 			mp = bp;
2976 		else
2977 			linkb(mp, bp);
2978 	} while (count > 0);
2979 
2980 	*mpp = mp;
2981 	return (0);
2982 }
2983 
2984 /*
2985  * Wait for a buffer to become available. Return non-zero errno
2986  * if not able to wait, 0 if buffer is probably there.
2987  */
2988 int
strwaitbuf(size_t size,int pri)2989 strwaitbuf(size_t size, int pri)
2990 {
2991 	bufcall_id_t id;
2992 
2993 	mutex_enter(&bcall_monitor);
2994 	if ((id = bufcall(size, pri, (void (*)(void *))cv_broadcast,
2995 	    &ttoproc(curthread)->p_flag_cv)) == 0) {
2996 		mutex_exit(&bcall_monitor);
2997 		return (ENOSR);
2998 	}
2999 	if (!cv_wait_sig(&(ttoproc(curthread)->p_flag_cv), &bcall_monitor)) {
3000 		unbufcall(id);
3001 		mutex_exit(&bcall_monitor);
3002 		return (EINTR);
3003 	}
3004 	unbufcall(id);
3005 	mutex_exit(&bcall_monitor);
3006 	return (0);
3007 }
3008 
3009 /*
3010  * This function waits for a read or write event to happen on a stream.
3011  * fmode can specify FNDELAY and/or FNONBLOCK.
3012  * The timeout is in ms with -1 meaning infinite.
3013  * The flag values work as follows:
3014  *	READWAIT	Check for read side errors, send M_READ
3015  *	GETWAIT		Check for read side errors, no M_READ
3016  *	WRITEWAIT	Check for write side errors.
3017  *	NOINTR		Do not return error if nonblocking or timeout.
3018  *	STR_NOERROR	Ignore all errors except STPLEX.
3019  *	STR_NOSIG	Ignore/hold signals during the duration of the call.
3020  *	STR_PEEK	Pass through the strgeterr().
3021  */
3022 int
strwaitq(stdata_t * stp,int flag,ssize_t count,int fmode,clock_t timout,int * done)3023 strwaitq(stdata_t *stp, int flag, ssize_t count, int fmode, clock_t timout,
3024     int *done)
3025 {
3026 	int slpflg, errs;
3027 	int error;
3028 	kcondvar_t *sleepon;
3029 	mblk_t *mp;
3030 	ssize_t *rd_count;
3031 	clock_t rval;
3032 
3033 	ASSERT(MUTEX_HELD(&stp->sd_lock));
3034 	if ((flag & READWAIT) || (flag & GETWAIT)) {
3035 		slpflg = RSLEEP;
3036 		sleepon = &_RD(stp->sd_wrq)->q_wait;
3037 		errs = STRDERR|STPLEX;
3038 	} else {
3039 		slpflg = WSLEEP;
3040 		sleepon = &stp->sd_wrq->q_wait;
3041 		errs = STWRERR|STRHUP|STPLEX;
3042 	}
3043 	if (flag & STR_NOERROR)
3044 		errs = STPLEX;
3045 
3046 	if (stp->sd_wakeq & slpflg) {
3047 		/*
3048 		 * A strwakeq() is pending, no need to sleep.
3049 		 */
3050 		stp->sd_wakeq &= ~slpflg;
3051 		*done = 0;
3052 		return (0);
3053 	}
3054 
3055 	if (stp->sd_flag & errs) {
3056 		/*
3057 		 * Check for errors before going to sleep since the
3058 		 * caller might not have checked this while holding
3059 		 * sd_lock.
3060 		 */
3061 		error = strgeterr(stp, errs, (flag & STR_PEEK));
3062 		if (error != 0) {
3063 			*done = 1;
3064 			return (error);
3065 		}
3066 	}
3067 
3068 	/*
3069 	 * If any module downstream has requested read notification
3070 	 * by setting SNDMREAD flag using M_SETOPTS, send a message
3071 	 * down stream.
3072 	 */
3073 	if ((flag & READWAIT) && (stp->sd_flag & SNDMREAD)) {
3074 		mutex_exit(&stp->sd_lock);
3075 		if (!(mp = allocb_wait(sizeof (ssize_t), BPRI_MED,
3076 		    (flag & STR_NOSIG), &error))) {
3077 			mutex_enter(&stp->sd_lock);
3078 			*done = 1;
3079 			return (error);
3080 		}
3081 		mp->b_datap->db_type = M_READ;
3082 		rd_count = (ssize_t *)mp->b_wptr;
3083 		*rd_count = count;
3084 		mp->b_wptr += sizeof (ssize_t);
3085 		/*
3086 		 * Send the number of bytes requested by the
3087 		 * read as the argument to M_READ.
3088 		 */
3089 		stream_willservice(stp);
3090 		putnext(stp->sd_wrq, mp);
3091 		stream_runservice(stp);
3092 		mutex_enter(&stp->sd_lock);
3093 
3094 		/*
3095 		 * If any data arrived due to inline processing
3096 		 * of putnext(), don't sleep.
3097 		 */
3098 		if (_RD(stp->sd_wrq)->q_first != NULL) {
3099 			*done = 0;
3100 			return (0);
3101 		}
3102 	}
3103 
3104 	if (fmode & (FNDELAY|FNONBLOCK)) {
3105 		if (!(flag & NOINTR))
3106 			error = EAGAIN;
3107 		else
3108 			error = 0;
3109 		*done = 1;
3110 		return (error);
3111 	}
3112 
3113 	stp->sd_flag |= slpflg;
3114 	TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_WAIT2,
3115 	    "strwaitq sleeps (2):%p, %X, %lX, %X, %p",
3116 	    stp, flag, count, fmode, done);
3117 
3118 	rval = str_cv_wait(sleepon, &stp->sd_lock, timout, flag & STR_NOSIG);
3119 	if (rval > 0) {
3120 		/* EMPTY */
3121 		TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_WAKE2,
3122 		    "strwaitq awakes(2):%X, %X, %X, %X, %X",
3123 		    stp, flag, count, fmode, done);
3124 	} else if (rval == 0) {
3125 		TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_INTR2,
3126 		    "strwaitq interrupt #2:%p, %X, %lX, %X, %p",
3127 		    stp, flag, count, fmode, done);
3128 		stp->sd_flag &= ~slpflg;
3129 		cv_broadcast(sleepon);
3130 		if (!(flag & NOINTR))
3131 			error = EINTR;
3132 		else
3133 			error = 0;
3134 		*done = 1;
3135 		return (error);
3136 	} else {
3137 		/* timeout */
3138 		TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_TIME,
3139 		    "strwaitq timeout:%p, %X, %lX, %X, %p",
3140 		    stp, flag, count, fmode, done);
3141 		*done = 1;
3142 		if (!(flag & NOINTR))
3143 			return (ETIME);
3144 		else
3145 			return (0);
3146 	}
3147 	/*
3148 	 * If the caller implements delayed errors (i.e. queued after data)
3149 	 * we can not check for errors here since data as well as an
3150 	 * error might have arrived at the stream head. We return to
3151 	 * have the caller check the read queue before checking for errors.
3152 	 */
3153 	if ((stp->sd_flag & errs) && !(flag & STR_DELAYERR)) {
3154 		error = strgeterr(stp, errs, (flag & STR_PEEK));
3155 		if (error != 0) {
3156 			*done = 1;
3157 			return (error);
3158 		}
3159 	}
3160 	*done = 0;
3161 	return (0);
3162 }
3163 
3164 /*
3165  * Perform job control discipline access checks.
3166  * Return 0 for success and the errno for failure.
3167  */
3168 
3169 #define	cantsend(p, t, sig) \
3170 	(sigismember(&(p)->p_ignore, sig) || signal_is_blocked((t), sig))
3171 
3172 int
straccess(struct stdata * stp,enum jcaccess mode)3173 straccess(struct stdata *stp, enum jcaccess mode)
3174 {
3175 	extern kcondvar_t lbolt_cv;	/* XXX: should be in a header file */
3176 	kthread_t *t = curthread;
3177 	proc_t *p = ttoproc(t);
3178 	sess_t *sp;
3179 
3180 	ASSERT(mutex_owned(&stp->sd_lock));
3181 
3182 	if (stp->sd_sidp == NULL || stp->sd_vnode->v_type == VFIFO)
3183 		return (0);
3184 
3185 	mutex_enter(&p->p_lock);		/* protects p_pgidp */
3186 
3187 	for (;;) {
3188 		mutex_enter(&p->p_splock);	/* protects p->p_sessp */
3189 		sp = p->p_sessp;
3190 		mutex_enter(&sp->s_lock);	/* protects sp->* */
3191 
3192 		/*
3193 		 * If this is not the calling process's controlling terminal
3194 		 * or if the calling process is already in the foreground
3195 		 * then allow access.
3196 		 */
3197 		if (sp->s_dev != stp->sd_vnode->v_rdev ||
3198 		    p->p_pgidp == stp->sd_pgidp) {
3199 			mutex_exit(&sp->s_lock);
3200 			mutex_exit(&p->p_splock);
3201 			mutex_exit(&p->p_lock);
3202 			return (0);
3203 		}
3204 
3205 		/*
3206 		 * Check to see if controlling terminal has been deallocated.
3207 		 */
3208 		if (sp->s_vp == NULL) {
3209 			if (!cantsend(p, t, SIGHUP))
3210 				sigtoproc(p, t, SIGHUP);
3211 			mutex_exit(&sp->s_lock);
3212 			mutex_exit(&p->p_splock);
3213 			mutex_exit(&p->p_lock);
3214 			return (EIO);
3215 		}
3216 
3217 		mutex_exit(&sp->s_lock);
3218 		mutex_exit(&p->p_splock);
3219 
3220 		if (mode == JCGETP) {
3221 			mutex_exit(&p->p_lock);
3222 			return (0);
3223 		}
3224 
3225 		if (mode == JCREAD) {
3226 			if (p->p_detached || cantsend(p, t, SIGTTIN)) {
3227 				mutex_exit(&p->p_lock);
3228 				return (EIO);
3229 			}
3230 			mutex_exit(&p->p_lock);
3231 			mutex_exit(&stp->sd_lock);
3232 			pgsignal(p->p_pgidp, SIGTTIN);
3233 			mutex_enter(&stp->sd_lock);
3234 			mutex_enter(&p->p_lock);
3235 		} else {  /* mode == JCWRITE or JCSETP */
3236 			if ((mode == JCWRITE && !(stp->sd_flag & STRTOSTOP)) ||
3237 			    cantsend(p, t, SIGTTOU)) {
3238 				mutex_exit(&p->p_lock);
3239 				return (0);
3240 			}
3241 			if (p->p_detached) {
3242 				mutex_exit(&p->p_lock);
3243 				return (EIO);
3244 			}
3245 			mutex_exit(&p->p_lock);
3246 			mutex_exit(&stp->sd_lock);
3247 			pgsignal(p->p_pgidp, SIGTTOU);
3248 			mutex_enter(&stp->sd_lock);
3249 			mutex_enter(&p->p_lock);
3250 		}
3251 
3252 		/*
3253 		 * We call cv_wait_sig_swap() to cause the appropriate
3254 		 * action for the jobcontrol signal to take place.
3255 		 * If the signal is being caught, we will take the
3256 		 * EINTR error return.  Otherwise, the default action
3257 		 * of causing the process to stop will take place.
3258 		 * In this case, we rely on the periodic cv_broadcast() on
3259 		 * &lbolt_cv to wake us up to loop around and test again.
3260 		 * We can't get here if the signal is ignored or
3261 		 * if the current thread is blocking the signal.
3262 		 */
3263 		mutex_exit(&stp->sd_lock);
3264 		if (!cv_wait_sig_swap(&lbolt_cv, &p->p_lock)) {
3265 			mutex_exit(&p->p_lock);
3266 			mutex_enter(&stp->sd_lock);
3267 			return (EINTR);
3268 		}
3269 		mutex_exit(&p->p_lock);
3270 		mutex_enter(&stp->sd_lock);
3271 		mutex_enter(&p->p_lock);
3272 	}
3273 }
3274 
3275 /*
3276  * Return size of message of block type (bp->b_datap->db_type)
3277  */
3278 size_t
xmsgsize(mblk_t * bp)3279 xmsgsize(mblk_t *bp)
3280 {
3281 	unsigned char type;
3282 	size_t count = 0;
3283 
3284 	type = bp->b_datap->db_type;
3285 
3286 	for (; bp; bp = bp->b_cont) {
3287 		if (type != bp->b_datap->db_type)
3288 			break;
3289 		ASSERT(bp->b_wptr >= bp->b_rptr);
3290 		count += bp->b_wptr - bp->b_rptr;
3291 	}
3292 	return (count);
3293 }
3294 
3295 /*
3296  * Allocate a stream head.
3297  */
3298 struct stdata *
shalloc(queue_t * qp)3299 shalloc(queue_t *qp)
3300 {
3301 	stdata_t *stp;
3302 
3303 	stp = kmem_cache_alloc(stream_head_cache, KM_SLEEP);
3304 
3305 	stp->sd_wrq = _WR(qp);
3306 	stp->sd_strtab = NULL;
3307 	stp->sd_iocid = 0;
3308 	stp->sd_mate = NULL;
3309 	stp->sd_freezer = NULL;
3310 	stp->sd_refcnt = 0;
3311 	stp->sd_wakeq = 0;
3312 	stp->sd_anchor = 0;
3313 	stp->sd_struiowrq = NULL;
3314 	stp->sd_struiordq = NULL;
3315 	stp->sd_struiodnak = 0;
3316 	stp->sd_struionak = NULL;
3317 	stp->sd_t_audit_data = NULL;
3318 	stp->sd_rput_opt = 0;
3319 	stp->sd_wput_opt = 0;
3320 	stp->sd_read_opt = 0;
3321 	stp->sd_rprotofunc = strrput_proto;
3322 	stp->sd_rmiscfunc = strrput_misc;
3323 	stp->sd_rderrfunc = stp->sd_wrerrfunc = NULL;
3324 	stp->sd_rputdatafunc = stp->sd_wputdatafunc = NULL;
3325 	stp->sd_ciputctrl = NULL;
3326 	stp->sd_nciputctrl = 0;
3327 	stp->sd_qhead = NULL;
3328 	stp->sd_qtail = NULL;
3329 	stp->sd_servid = NULL;
3330 	stp->sd_nqueues = 0;
3331 	stp->sd_svcflags = 0;
3332 	stp->sd_copyflag = 0;
3333 
3334 	return (stp);
3335 }
3336 
3337 /*
3338  * Free a stream head.
3339  */
3340 void
shfree(stdata_t * stp)3341 shfree(stdata_t *stp)
3342 {
3343 	ASSERT(MUTEX_NOT_HELD(&stp->sd_lock));
3344 
3345 	stp->sd_wrq = NULL;
3346 
3347 	mutex_enter(&stp->sd_qlock);
3348 	while (stp->sd_svcflags & STRS_SCHEDULED) {
3349 		STRSTAT(strwaits);
3350 		cv_wait(&stp->sd_qcv, &stp->sd_qlock);
3351 	}
3352 	mutex_exit(&stp->sd_qlock);
3353 
3354 	if (stp->sd_ciputctrl != NULL) {
3355 		ASSERT(stp->sd_nciputctrl == n_ciputctrl - 1);
3356 		SUMCHECK_CIPUTCTRL_COUNTS(stp->sd_ciputctrl,
3357 		    stp->sd_nciputctrl, 0);
3358 		ASSERT(ciputctrl_cache != NULL);
3359 		kmem_cache_free(ciputctrl_cache, stp->sd_ciputctrl);
3360 		stp->sd_ciputctrl = NULL;
3361 		stp->sd_nciputctrl = 0;
3362 	}
3363 	ASSERT(stp->sd_qhead == NULL);
3364 	ASSERT(stp->sd_qtail == NULL);
3365 	ASSERT(stp->sd_nqueues == 0);
3366 	kmem_cache_free(stream_head_cache, stp);
3367 }
3368 
3369 /*
3370  * Allocate a pair of queues and a syncq for the pair
3371  */
3372 queue_t *
allocq(void)3373 allocq(void)
3374 {
3375 	queinfo_t *qip;
3376 	queue_t *qp, *wqp;
3377 	syncq_t	*sq;
3378 
3379 	qip = kmem_cache_alloc(queue_cache, KM_SLEEP);
3380 
3381 	qp = &qip->qu_rqueue;
3382 	wqp = &qip->qu_wqueue;
3383 	sq = &qip->qu_syncq;
3384 
3385 	qp->q_last	= NULL;
3386 	qp->q_next	= NULL;
3387 	qp->q_ptr	= NULL;
3388 	qp->q_flag	= QUSE | QREADR;
3389 	qp->q_bandp	= NULL;
3390 	qp->q_stream	= NULL;
3391 	qp->q_syncq	= sq;
3392 	qp->q_nband	= 0;
3393 	qp->q_nfsrv	= NULL;
3394 	qp->q_draining	= 0;
3395 	qp->q_syncqmsgs	= 0;
3396 	qp->q_spri	= 0;
3397 	qp->q_qtstamp	= 0;
3398 	qp->q_sqtstamp	= 0;
3399 	qp->q_fp	= NULL;
3400 
3401 	wqp->q_last	= NULL;
3402 	wqp->q_next	= NULL;
3403 	wqp->q_ptr	= NULL;
3404 	wqp->q_flag	= QUSE;
3405 	wqp->q_bandp	= NULL;
3406 	wqp->q_stream	= NULL;
3407 	wqp->q_syncq	= sq;
3408 	wqp->q_nband	= 0;
3409 	wqp->q_nfsrv	= NULL;
3410 	wqp->q_draining	= 0;
3411 	wqp->q_syncqmsgs = 0;
3412 	wqp->q_qtstamp	= 0;
3413 	wqp->q_sqtstamp	= 0;
3414 	wqp->q_spri	= 0;
3415 
3416 	sq->sq_count	= 0;
3417 	sq->sq_rmqcount	= 0;
3418 	sq->sq_flags	= 0;
3419 	sq->sq_type	= 0;
3420 	sq->sq_callbflags = 0;
3421 	sq->sq_cancelid	= 0;
3422 	sq->sq_ciputctrl = NULL;
3423 	sq->sq_nciputctrl = 0;
3424 	sq->sq_needexcl = 0;
3425 	sq->sq_svcflags = 0;
3426 
3427 	return (qp);
3428 }
3429 
3430 /*
3431  * Free a pair of queues and the "attached" syncq.
3432  * Discard any messages left on the syncq(s), remove the syncq(s) from the
3433  * outer perimeter, and free the syncq(s) if they are not the "attached" syncq.
3434  */
3435 void
freeq(queue_t * qp)3436 freeq(queue_t *qp)
3437 {
3438 	qband_t *qbp, *nqbp;
3439 	syncq_t *sq, *outer;
3440 	queue_t *wqp = _WR(qp);
3441 
3442 	ASSERT(qp->q_flag & QREADR);
3443 
3444 	/*
3445 	 * If a previously dispatched taskq job is scheduled to run
3446 	 * sync_service() or a service routine is scheduled for the
3447 	 * queues about to be freed, wait here until all service is
3448 	 * done on the queue and all associated queues and syncqs.
3449 	 */
3450 	wait_svc(qp);
3451 
3452 	(void) flush_syncq(qp->q_syncq, qp);
3453 	(void) flush_syncq(wqp->q_syncq, wqp);
3454 	ASSERT(qp->q_syncqmsgs == 0 && wqp->q_syncqmsgs == 0);
3455 
3456 	/*
3457 	 * Flush the queues before q_next is set to NULL This is needed
3458 	 * in order to backenable any downstream queue before we go away.
3459 	 * Note: we are already removed from the stream so that the
3460 	 * backenabling will not cause any messages to be delivered to our
3461 	 * put procedures.
3462 	 */
3463 	flushq(qp, FLUSHALL);
3464 	flushq(wqp, FLUSHALL);
3465 
3466 	/* Tidy up - removeq only does a half-remove from stream */
3467 	qp->q_next = wqp->q_next = NULL;
3468 	ASSERT(!(qp->q_flag & QENAB));
3469 	ASSERT(!(wqp->q_flag & QENAB));
3470 
3471 	outer = qp->q_syncq->sq_outer;
3472 	if (outer != NULL) {
3473 		outer_remove(outer, qp->q_syncq);
3474 		if (wqp->q_syncq != qp->q_syncq)
3475 			outer_remove(outer, wqp->q_syncq);
3476 	}
3477 	/*
3478 	 * Free any syncqs that are outside what allocq returned.
3479 	 */
3480 	if (qp->q_syncq != SQ(qp) && !(qp->q_flag & QPERMOD))
3481 		free_syncq(qp->q_syncq);
3482 	if (qp->q_syncq != wqp->q_syncq && wqp->q_syncq != SQ(qp))
3483 		free_syncq(wqp->q_syncq);
3484 
3485 	ASSERT((qp->q_sqflags & (Q_SQQUEUED | Q_SQDRAINING)) == 0);
3486 	ASSERT((wqp->q_sqflags & (Q_SQQUEUED | Q_SQDRAINING)) == 0);
3487 	ASSERT(MUTEX_NOT_HELD(QLOCK(qp)));
3488 	ASSERT(MUTEX_NOT_HELD(QLOCK(wqp)));
3489 	sq = SQ(qp);
3490 	ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
3491 	ASSERT(sq->sq_head == NULL && sq->sq_tail == NULL);
3492 	ASSERT(sq->sq_outer == NULL);
3493 	ASSERT(sq->sq_onext == NULL && sq->sq_oprev == NULL);
3494 	ASSERT(sq->sq_callbpend == NULL);
3495 	ASSERT(sq->sq_needexcl == 0);
3496 
3497 	if (sq->sq_ciputctrl != NULL) {
3498 		ASSERT(sq->sq_nciputctrl == n_ciputctrl - 1);
3499 		SUMCHECK_CIPUTCTRL_COUNTS(sq->sq_ciputctrl,
3500 		    sq->sq_nciputctrl, 0);
3501 		ASSERT(ciputctrl_cache != NULL);
3502 		kmem_cache_free(ciputctrl_cache, sq->sq_ciputctrl);
3503 		sq->sq_ciputctrl = NULL;
3504 		sq->sq_nciputctrl = 0;
3505 	}
3506 
3507 	ASSERT(qp->q_first == NULL && wqp->q_first == NULL);
3508 	ASSERT(qp->q_count == 0 && wqp->q_count == 0);
3509 	ASSERT(qp->q_mblkcnt == 0 && wqp->q_mblkcnt == 0);
3510 
3511 	qp->q_flag &= ~QUSE;
3512 	wqp->q_flag &= ~QUSE;
3513 
3514 	/* NOTE: Uncomment the assert below once bugid 1159635 is fixed. */
3515 	/* ASSERT((qp->q_flag & QWANTW) == 0 && (wqp->q_flag & QWANTW) == 0); */
3516 
3517 	qbp = qp->q_bandp;
3518 	while (qbp) {
3519 		nqbp = qbp->qb_next;
3520 		freeband(qbp);
3521 		qbp = nqbp;
3522 	}
3523 	qbp = wqp->q_bandp;
3524 	while (qbp) {
3525 		nqbp = qbp->qb_next;
3526 		freeband(qbp);
3527 		qbp = nqbp;
3528 	}
3529 	kmem_cache_free(queue_cache, qp);
3530 }
3531 
3532 /*
3533  * Allocate a qband structure.
3534  */
3535 qband_t *
allocband(void)3536 allocband(void)
3537 {
3538 	qband_t *qbp;
3539 
3540 	qbp = kmem_cache_alloc(qband_cache, KM_NOSLEEP);
3541 	if (qbp == NULL)
3542 		return (NULL);
3543 
3544 	qbp->qb_next	= NULL;
3545 	qbp->qb_count	= 0;
3546 	qbp->qb_mblkcnt	= 0;
3547 	qbp->qb_first	= NULL;
3548 	qbp->qb_last	= NULL;
3549 	qbp->qb_flag	= 0;
3550 
3551 	return (qbp);
3552 }
3553 
3554 /*
3555  * Free a qband structure.
3556  */
3557 void
freeband(qband_t * qbp)3558 freeband(qband_t *qbp)
3559 {
3560 	kmem_cache_free(qband_cache, qbp);
3561 }
3562 
3563 /*
3564  * Just like putnextctl(9F), except that allocb_wait() is used.
3565  *
3566  * Consolidation Private, and of course only callable from the stream head or
3567  * routines that may block.
3568  */
3569 int
putnextctl_wait(queue_t * q,int type)3570 putnextctl_wait(queue_t *q, int type)
3571 {
3572 	mblk_t *bp;
3573 	int error;
3574 
3575 	if ((datamsg(type) && (type != M_DELAY)) ||
3576 	    (bp = allocb_wait(0, BPRI_HI, 0, &error)) == NULL)
3577 		return (0);
3578 
3579 	bp->b_datap->db_type = (unsigned char)type;
3580 	putnext(q, bp);
3581 	return (1);
3582 }
3583 
3584 /*
3585  * Run any possible bufcalls.
3586  */
3587 void
runbufcalls(void)3588 runbufcalls(void)
3589 {
3590 	strbufcall_t *bcp;
3591 
3592 	mutex_enter(&bcall_monitor);
3593 	mutex_enter(&strbcall_lock);
3594 
3595 	if (strbcalls.bc_head) {
3596 		size_t count;
3597 		int nevent;
3598 
3599 		/*
3600 		 * count how many events are on the list
3601 		 * now so we can check to avoid looping
3602 		 * in low memory situations
3603 		 */
3604 		nevent = 0;
3605 		for (bcp = strbcalls.bc_head; bcp; bcp = bcp->bc_next)
3606 			nevent++;
3607 
3608 		/*
3609 		 * get estimate of available memory from kmem_avail().
3610 		 * awake all bufcall functions waiting for
3611 		 * memory whose request could be satisfied
3612 		 * by 'count' memory and let 'em fight for it.
3613 		 */
3614 		count = kmem_avail();
3615 		while ((bcp = strbcalls.bc_head) != NULL && nevent) {
3616 			STRSTAT(bufcalls);
3617 			--nevent;
3618 			if (bcp->bc_size <= count) {
3619 				bcp->bc_executor = curthread;
3620 				mutex_exit(&strbcall_lock);
3621 				(*bcp->bc_func)(bcp->bc_arg);
3622 				mutex_enter(&strbcall_lock);
3623 				bcp->bc_executor = NULL;
3624 				cv_broadcast(&bcall_cv);
3625 				strbcalls.bc_head = bcp->bc_next;
3626 				kmem_free(bcp, sizeof (strbufcall_t));
3627 			} else {
3628 				/*
3629 				 * too big, try again later - note
3630 				 * that nevent was decremented above
3631 				 * so we won't retry this one on this
3632 				 * iteration of the loop
3633 				 */
3634 				if (bcp->bc_next != NULL) {
3635 					strbcalls.bc_head = bcp->bc_next;
3636 					bcp->bc_next = NULL;
3637 					strbcalls.bc_tail->bc_next = bcp;
3638 					strbcalls.bc_tail = bcp;
3639 				}
3640 			}
3641 		}
3642 		if (strbcalls.bc_head == NULL)
3643 			strbcalls.bc_tail = NULL;
3644 	}
3645 
3646 	mutex_exit(&strbcall_lock);
3647 	mutex_exit(&bcall_monitor);
3648 }
3649 
3650 
3651 /*
3652  * Actually run queue's service routine.
3653  */
3654 static void
runservice(queue_t * q)3655 runservice(queue_t *q)
3656 {
3657 	qband_t *qbp;
3658 
3659 	ASSERT(q->q_qinfo->qi_srvp);
3660 again:
3661 	entersq(q->q_syncq, SQ_SVC);
3662 	TRACE_1(TR_FAC_STREAMS_FR, TR_QRUNSERVICE_START,
3663 	    "runservice starts:%p", q);
3664 
3665 	if (!(q->q_flag & QWCLOSE))
3666 		(*q->q_qinfo->qi_srvp)(q);
3667 
3668 	TRACE_1(TR_FAC_STREAMS_FR, TR_QRUNSERVICE_END,
3669 	    "runservice ends:(%p)", q);
3670 
3671 	leavesq(q->q_syncq, SQ_SVC);
3672 
3673 	mutex_enter(QLOCK(q));
3674 	if (q->q_flag & QENAB) {
3675 		q->q_flag &= ~QENAB;
3676 		mutex_exit(QLOCK(q));
3677 		goto again;
3678 	}
3679 	q->q_flag &= ~QINSERVICE;
3680 	q->q_flag &= ~QBACK;
3681 	for (qbp = q->q_bandp; qbp; qbp = qbp->qb_next)
3682 		qbp->qb_flag &= ~QB_BACK;
3683 	/*
3684 	 * Wakeup thread waiting for the service procedure
3685 	 * to be run (strclose and qdetach).
3686 	 */
3687 	cv_broadcast(&q->q_wait);
3688 
3689 	mutex_exit(QLOCK(q));
3690 }
3691 
3692 /*
3693  * Background processing of bufcalls.
3694  */
3695 void
streams_bufcall_service(void)3696 streams_bufcall_service(void)
3697 {
3698 	callb_cpr_t	cprinfo;
3699 
3700 	CALLB_CPR_INIT(&cprinfo, &strbcall_lock, callb_generic_cpr,
3701 	    "streams_bufcall_service");
3702 
3703 	mutex_enter(&strbcall_lock);
3704 
3705 	for (;;) {
3706 		if (strbcalls.bc_head != NULL && kmem_avail() > 0) {
3707 			mutex_exit(&strbcall_lock);
3708 			runbufcalls();
3709 			mutex_enter(&strbcall_lock);
3710 		}
3711 		if (strbcalls.bc_head != NULL) {
3712 			STRSTAT(bcwaits);
3713 			/* Wait for memory to become available */
3714 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3715 			(void) cv_reltimedwait(&memavail_cv, &strbcall_lock,
3716 			    SEC_TO_TICK(60), TR_CLOCK_TICK);
3717 			CALLB_CPR_SAFE_END(&cprinfo, &strbcall_lock);
3718 		}
3719 
3720 		/* Wait for new work to arrive */
3721 		if (strbcalls.bc_head == NULL) {
3722 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3723 			cv_wait(&strbcall_cv, &strbcall_lock);
3724 			CALLB_CPR_SAFE_END(&cprinfo, &strbcall_lock);
3725 		}
3726 	}
3727 }
3728 
3729 /*
3730  * Background processing of streams background tasks which failed
3731  * taskq_dispatch.
3732  */
3733 static void
streams_qbkgrnd_service(void)3734 streams_qbkgrnd_service(void)
3735 {
3736 	callb_cpr_t cprinfo;
3737 	queue_t *q;
3738 
3739 	CALLB_CPR_INIT(&cprinfo, &service_queue, callb_generic_cpr,
3740 	    "streams_bkgrnd_service");
3741 
3742 	mutex_enter(&service_queue);
3743 
3744 	for (;;) {
3745 		/*
3746 		 * Wait for work to arrive.
3747 		 */
3748 		while ((freebs_list == NULL) && (qhead == NULL)) {
3749 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3750 			cv_wait(&services_to_run, &service_queue);
3751 			CALLB_CPR_SAFE_END(&cprinfo, &service_queue);
3752 		}
3753 		/*
3754 		 * Handle all pending freebs requests to free memory.
3755 		 */
3756 		while (freebs_list != NULL) {
3757 			mblk_t *mp = freebs_list;
3758 			freebs_list = mp->b_next;
3759 			mutex_exit(&service_queue);
3760 			mblk_free(mp);
3761 			mutex_enter(&service_queue);
3762 		}
3763 		/*
3764 		 * Run pending queues.
3765 		 */
3766 		while (qhead != NULL) {
3767 			DQ(q, qhead, qtail, q_link);
3768 			ASSERT(q != NULL);
3769 			mutex_exit(&service_queue);
3770 			queue_service(q);
3771 			mutex_enter(&service_queue);
3772 		}
3773 		ASSERT(qhead == NULL && qtail == NULL);
3774 	}
3775 }
3776 
3777 /*
3778  * Background processing of streams background tasks which failed
3779  * taskq_dispatch.
3780  */
3781 static void
streams_sqbkgrnd_service(void)3782 streams_sqbkgrnd_service(void)
3783 {
3784 	callb_cpr_t cprinfo;
3785 	syncq_t *sq;
3786 
3787 	CALLB_CPR_INIT(&cprinfo, &service_queue, callb_generic_cpr,
3788 	    "streams_sqbkgrnd_service");
3789 
3790 	mutex_enter(&service_queue);
3791 
3792 	for (;;) {
3793 		/*
3794 		 * Wait for work to arrive.
3795 		 */
3796 		while (sqhead == NULL) {
3797 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3798 			cv_wait(&syncqs_to_run, &service_queue);
3799 			CALLB_CPR_SAFE_END(&cprinfo, &service_queue);
3800 		}
3801 
3802 		/*
3803 		 * Run pending syncqs.
3804 		 */
3805 		while (sqhead != NULL) {
3806 			DQ(sq, sqhead, sqtail, sq_next);
3807 			ASSERT(sq != NULL);
3808 			ASSERT(sq->sq_svcflags & SQ_BGTHREAD);
3809 			mutex_exit(&service_queue);
3810 			syncq_service(sq);
3811 			mutex_enter(&service_queue);
3812 		}
3813 	}
3814 }
3815 
3816 /*
3817  * Disable the syncq and wait for background syncq processing to complete.
3818  * If the syncq is placed on the sqhead/sqtail queue, try to remove it from the
3819  * list.
3820  */
3821 void
wait_sq_svc(syncq_t * sq)3822 wait_sq_svc(syncq_t *sq)
3823 {
3824 	mutex_enter(SQLOCK(sq));
3825 	sq->sq_svcflags |= SQ_DISABLED;
3826 	if (sq->sq_svcflags & SQ_BGTHREAD) {
3827 		syncq_t *sq_chase;
3828 		syncq_t *sq_curr;
3829 		int removed;
3830 
3831 		ASSERT(sq->sq_servcount == 1);
3832 		mutex_enter(&service_queue);
3833 		RMQ(sq, sqhead, sqtail, sq_next, sq_chase, sq_curr, removed);
3834 		mutex_exit(&service_queue);
3835 		if (removed) {
3836 			sq->sq_svcflags &= ~SQ_BGTHREAD;
3837 			sq->sq_servcount = 0;
3838 			STRSTAT(sqremoved);
3839 			goto done;
3840 		}
3841 	}
3842 	while (sq->sq_servcount != 0) {
3843 		sq->sq_flags |= SQ_WANTWAKEUP;
3844 		cv_wait(&sq->sq_wait, SQLOCK(sq));
3845 	}
3846 done:
3847 	mutex_exit(SQLOCK(sq));
3848 }
3849 
3850 /*
3851  * Put a syncq on the list of syncq's to be serviced by the sqthread.
3852  * Add the argument to the end of the sqhead list and set the flag
3853  * indicating this syncq has been enabled.  If it has already been
3854  * enabled, don't do anything.
3855  * This routine assumes that SQLOCK is held.
3856  * NOTE that the lock order is to have the SQLOCK first,
3857  * so if the service_syncq lock is held, we need to release it
3858  * before acquiring the SQLOCK (mostly relevant for the background
3859  * thread, and this seems to be common among the STREAMS global locks).
3860  * Note that the sq_svcflags are protected by the SQLOCK.
3861  */
3862 void
sqenable(syncq_t * sq)3863 sqenable(syncq_t *sq)
3864 {
3865 	/*
3866 	 * This is probably not important except for where I believe it
3867 	 * is being called.  At that point, it should be held (and it
3868 	 * is a pain to release it just for this routine, so don't do
3869 	 * it).
3870 	 */
3871 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
3872 
3873 	IMPLY(sq->sq_servcount == 0, sq->sq_next == NULL);
3874 	IMPLY(sq->sq_next != NULL, sq->sq_svcflags & SQ_BGTHREAD);
3875 
3876 	/*
3877 	 * Do not put on list if background thread is scheduled or
3878 	 * syncq is disabled.
3879 	 */
3880 	if (sq->sq_svcflags & (SQ_DISABLED | SQ_BGTHREAD))
3881 		return;
3882 
3883 	/*
3884 	 * Check whether we should enable sq at all.
3885 	 * Non PERMOD syncqs may be drained by at most one thread.
3886 	 * PERMOD syncqs may be drained by several threads but we limit the
3887 	 * total amount to the lesser of
3888 	 *	Number of queues on the squeue and
3889 	 *	Number of CPUs.
3890 	 */
3891 	if (sq->sq_servcount != 0) {
3892 		if (((sq->sq_type & SQ_PERMOD) == 0) ||
3893 		    (sq->sq_servcount >= MIN(sq->sq_nqueues, ncpus_online))) {
3894 			STRSTAT(sqtoomany);
3895 			return;
3896 		}
3897 	}
3898 
3899 	sq->sq_tstamp = ddi_get_lbolt();
3900 	STRSTAT(sqenables);
3901 
3902 	/* Attempt a taskq dispatch */
3903 	sq->sq_servid = (void *)taskq_dispatch(streams_taskq,
3904 	    (task_func_t *)syncq_service, sq, TQ_NOSLEEP | TQ_NOQUEUE);
3905 	if (sq->sq_servid != NULL) {
3906 		sq->sq_servcount++;
3907 		return;
3908 	}
3909 
3910 	/*
3911 	 * This taskq dispatch failed, but a previous one may have succeeded.
3912 	 * Don't try to schedule on the background thread whilst there is
3913 	 * outstanding taskq processing.
3914 	 */
3915 	if (sq->sq_servcount != 0)
3916 		return;
3917 
3918 	/*
3919 	 * System is low on resources and can't perform a non-sleeping
3920 	 * dispatch. Schedule the syncq for a background thread and mark the
3921 	 * syncq to avoid any further taskq dispatch attempts.
3922 	 */
3923 	mutex_enter(&service_queue);
3924 	STRSTAT(taskqfails);
3925 	ENQUEUE(sq, sqhead, sqtail, sq_next);
3926 	sq->sq_svcflags |= SQ_BGTHREAD;
3927 	sq->sq_servcount = 1;
3928 	cv_signal(&syncqs_to_run);
3929 	mutex_exit(&service_queue);
3930 }
3931 
3932 /*
3933  * Note: fifo_close() depends on the mblk_t on the queue being freed
3934  * asynchronously. The asynchronous freeing of messages breaks the
3935  * recursive call chain of fifo_close() while there are I_SENDFD type of
3936  * messages referring to other file pointers on the queue. Then when
3937  * closing pipes it can avoid stack overflow in case of daisy-chained
3938  * pipes, and also avoid deadlock in case of fifonode_t pairs (which
3939  * share the same fifolock_t).
3940  *
3941  * No need to kpreempt_disable to access cpu_seqid.  If we migrate and
3942  * the esb queue does not match the new CPU, that is OK.
3943  */
3944 void
freebs_enqueue(mblk_t * mp,dblk_t * dbp)3945 freebs_enqueue(mblk_t *mp, dblk_t *dbp)
3946 {
3947 	int qindex = CPU->cpu_seqid >> esbq_log2_cpus_per_q;
3948 	esb_queue_t *eqp;
3949 
3950 	ASSERT(dbp->db_mblk == mp);
3951 	ASSERT(qindex < esbq_nelem);
3952 
3953 	eqp = system_esbq_array;
3954 	if (eqp != NULL) {
3955 		eqp += qindex;
3956 	} else {
3957 		mutex_enter(&esbq_lock);
3958 		if (kmem_ready && system_esbq_array == NULL)
3959 			system_esbq_array = (esb_queue_t *)kmem_zalloc(
3960 			    esbq_nelem * sizeof (esb_queue_t), KM_NOSLEEP);
3961 		mutex_exit(&esbq_lock);
3962 		eqp = system_esbq_array;
3963 		if (eqp != NULL)
3964 			eqp += qindex;
3965 		else
3966 			eqp = &system_esbq;
3967 	}
3968 
3969 	/*
3970 	 * Check data sanity. The dblock should have non-empty free function.
3971 	 * It is better to panic here then later when the dblock is freed
3972 	 * asynchronously when the context is lost.
3973 	 */
3974 	if (dbp->db_frtnp->free_func == NULL) {
3975 		panic("freebs_enqueue: dblock %p has a NULL free callback",
3976 		    (void *)dbp);
3977 	}
3978 
3979 	mutex_enter(&eqp->eq_lock);
3980 	/* queue the new mblk on the esballoc queue */
3981 	if (eqp->eq_head == NULL) {
3982 		eqp->eq_head = eqp->eq_tail = mp;
3983 	} else {
3984 		eqp->eq_tail->b_next = mp;
3985 		eqp->eq_tail = mp;
3986 	}
3987 	eqp->eq_len++;
3988 
3989 	/* If we're the first thread to reach the threshold, process */
3990 	if (eqp->eq_len >= esbq_max_qlen &&
3991 	    !(eqp->eq_flags & ESBQ_PROCESSING))
3992 		esballoc_process_queue(eqp);
3993 
3994 	esballoc_set_timer(eqp, esbq_timeout);
3995 	mutex_exit(&eqp->eq_lock);
3996 }
3997 
3998 static void
esballoc_process_queue(esb_queue_t * eqp)3999 esballoc_process_queue(esb_queue_t *eqp)
4000 {
4001 	mblk_t	*mp;
4002 
4003 	ASSERT(MUTEX_HELD(&eqp->eq_lock));
4004 
4005 	eqp->eq_flags |= ESBQ_PROCESSING;
4006 
4007 	do {
4008 		/*
4009 		 * Detach the message chain for processing.
4010 		 */
4011 		mp = eqp->eq_head;
4012 		eqp->eq_tail->b_next = NULL;
4013 		eqp->eq_head = eqp->eq_tail = NULL;
4014 		eqp->eq_len = 0;
4015 		mutex_exit(&eqp->eq_lock);
4016 
4017 		/*
4018 		 * Process the message chain.
4019 		 */
4020 		esballoc_enqueue_mblk(mp);
4021 		mutex_enter(&eqp->eq_lock);
4022 	} while ((eqp->eq_len >= esbq_max_qlen) && (eqp->eq_len > 0));
4023 
4024 	eqp->eq_flags &= ~ESBQ_PROCESSING;
4025 }
4026 
4027 /*
4028  * taskq callback routine to free esballoced mblk's
4029  */
4030 static void
esballoc_mblk_free(mblk_t * mp)4031 esballoc_mblk_free(mblk_t *mp)
4032 {
4033 	mblk_t	*nextmp;
4034 
4035 	for (; mp != NULL; mp = nextmp) {
4036 		nextmp = mp->b_next;
4037 		mp->b_next = NULL;
4038 		mblk_free(mp);
4039 	}
4040 }
4041 
4042 static void
esballoc_enqueue_mblk(mblk_t * mp)4043 esballoc_enqueue_mblk(mblk_t *mp)
4044 {
4045 
4046 	if (taskq_dispatch(system_taskq, (task_func_t *)esballoc_mblk_free, mp,
4047 	    TQ_NOSLEEP) == TASKQID_INVALID) {
4048 		mblk_t *first_mp = mp;
4049 		/*
4050 		 * System is low on resources and can't perform a non-sleeping
4051 		 * dispatch. Schedule for a background thread.
4052 		 */
4053 		mutex_enter(&service_queue);
4054 		STRSTAT(taskqfails);
4055 
4056 		while (mp->b_next != NULL)
4057 			mp = mp->b_next;
4058 
4059 		mp->b_next = freebs_list;
4060 		freebs_list = first_mp;
4061 		cv_signal(&services_to_run);
4062 		mutex_exit(&service_queue);
4063 	}
4064 }
4065 
4066 static void
esballoc_timer(void * arg)4067 esballoc_timer(void *arg)
4068 {
4069 	esb_queue_t *eqp = arg;
4070 
4071 	mutex_enter(&eqp->eq_lock);
4072 	eqp->eq_flags &= ~ESBQ_TIMER;
4073 
4074 	if (!(eqp->eq_flags & ESBQ_PROCESSING) &&
4075 	    eqp->eq_len > 0)
4076 		esballoc_process_queue(eqp);
4077 
4078 	esballoc_set_timer(eqp, esbq_timeout);
4079 	mutex_exit(&eqp->eq_lock);
4080 }
4081 
4082 static void
esballoc_set_timer(esb_queue_t * eqp,clock_t eq_timeout)4083 esballoc_set_timer(esb_queue_t *eqp, clock_t eq_timeout)
4084 {
4085 	ASSERT(MUTEX_HELD(&eqp->eq_lock));
4086 
4087 	if (eqp->eq_len > 0 && !(eqp->eq_flags & ESBQ_TIMER)) {
4088 		(void) timeout(esballoc_timer, eqp, eq_timeout);
4089 		eqp->eq_flags |= ESBQ_TIMER;
4090 	}
4091 }
4092 
4093 /*
4094  * Setup esbq array length based upon NCPU scaled by CPUs per
4095  * queue. Use static system_esbq until kmem_ready and we can
4096  * create an array in freebs_enqueue().
4097  */
4098 void
esballoc_queue_init(void)4099 esballoc_queue_init(void)
4100 {
4101 	esbq_log2_cpus_per_q = highbit(esbq_cpus_per_q - 1);
4102 	esbq_cpus_per_q = 1 << esbq_log2_cpus_per_q;
4103 	esbq_nelem = howmany(NCPU, esbq_cpus_per_q);
4104 	system_esbq.eq_len = 0;
4105 	system_esbq.eq_head = system_esbq.eq_tail = NULL;
4106 	system_esbq.eq_flags = 0;
4107 }
4108 
4109 /*
4110  * Set the QBACK or QB_BACK flag in the given queue for
4111  * the given priority band.
4112  */
4113 void
setqback(queue_t * q,unsigned char pri)4114 setqback(queue_t *q, unsigned char pri)
4115 {
4116 	int i;
4117 	qband_t *qbp;
4118 	qband_t **qbpp;
4119 
4120 	ASSERT(MUTEX_HELD(QLOCK(q)));
4121 	if (pri != 0) {
4122 		if (pri > q->q_nband) {
4123 			qbpp = &q->q_bandp;
4124 			while (*qbpp)
4125 				qbpp = &(*qbpp)->qb_next;
4126 			while (pri > q->q_nband) {
4127 				if ((*qbpp = allocband()) == NULL) {
4128 					cmn_err(CE_WARN,
4129 					    "setqback: can't allocate qband\n");
4130 					return;
4131 				}
4132 				(*qbpp)->qb_hiwat = q->q_hiwat;
4133 				(*qbpp)->qb_lowat = q->q_lowat;
4134 				q->q_nband++;
4135 				qbpp = &(*qbpp)->qb_next;
4136 			}
4137 		}
4138 		qbp = q->q_bandp;
4139 		i = pri;
4140 		while (--i)
4141 			qbp = qbp->qb_next;
4142 		qbp->qb_flag |= QB_BACK;
4143 	} else {
4144 		q->q_flag |= QBACK;
4145 	}
4146 }
4147 
4148 int
strcopyin(void * from,void * to,size_t len,int copyflag)4149 strcopyin(void *from, void *to, size_t len, int copyflag)
4150 {
4151 	if (copyflag & U_TO_K) {
4152 		ASSERT((copyflag & K_TO_K) == 0);
4153 		if (copyin(from, to, len))
4154 			return (EFAULT);
4155 	} else {
4156 		ASSERT(copyflag & K_TO_K);
4157 		bcopy(from, to, len);
4158 	}
4159 	return (0);
4160 }
4161 
4162 int
strcopyout(void * from,void * to,size_t len,int copyflag)4163 strcopyout(void *from, void *to, size_t len, int copyflag)
4164 {
4165 	if (copyflag & U_TO_K) {
4166 		if (copyout(from, to, len))
4167 			return (EFAULT);
4168 	} else {
4169 		ASSERT(copyflag & K_TO_K);
4170 		bcopy(from, to, len);
4171 	}
4172 	return (0);
4173 }
4174 
4175 /*
4176  * strsignal_nolock() posts a signal to the process(es) at the stream head.
4177  * It assumes that the stream head lock is already held, whereas strsignal()
4178  * acquires the lock first.  This routine was created because a few callers
4179  * release the stream head lock before calling only to re-acquire it after
4180  * it returns.
4181  */
4182 void
strsignal_nolock(stdata_t * stp,int sig,uchar_t band)4183 strsignal_nolock(stdata_t *stp, int sig, uchar_t band)
4184 {
4185 	ASSERT(MUTEX_HELD(&stp->sd_lock));
4186 	switch (sig) {
4187 	case SIGPOLL:
4188 		if (stp->sd_sigflags & S_MSG)
4189 			strsendsig(stp->sd_siglist, S_MSG, band, 0);
4190 		break;
4191 	default:
4192 		if (stp->sd_pgidp)
4193 			pgsignal(stp->sd_pgidp, sig);
4194 		break;
4195 	}
4196 }
4197 
4198 void
strsignal(stdata_t * stp,int sig,int32_t band)4199 strsignal(stdata_t *stp, int sig, int32_t band)
4200 {
4201 	TRACE_3(TR_FAC_STREAMS_FR, TR_SENDSIG,
4202 	    "strsignal:%p, %X, %X", stp, sig, band);
4203 
4204 	mutex_enter(&stp->sd_lock);
4205 	switch (sig) {
4206 	case SIGPOLL:
4207 		if (stp->sd_sigflags & S_MSG)
4208 			strsendsig(stp->sd_siglist, S_MSG, (uchar_t)band, 0);
4209 		break;
4210 
4211 	default:
4212 		if (stp->sd_pgidp) {
4213 			pgsignal(stp->sd_pgidp, sig);
4214 		}
4215 		break;
4216 	}
4217 	mutex_exit(&stp->sd_lock);
4218 }
4219 
4220 void
strhup(stdata_t * stp)4221 strhup(stdata_t *stp)
4222 {
4223 	ASSERT(mutex_owned(&stp->sd_lock));
4224 	pollwakeup(&stp->sd_pollist, POLLHUP);
4225 	if (stp->sd_sigflags & S_HANGUP)
4226 		strsendsig(stp->sd_siglist, S_HANGUP, 0, 0);
4227 }
4228 
4229 /*
4230  * Backenable the first queue upstream from `q' with a service procedure.
4231  */
4232 void
backenable(queue_t * q,uchar_t pri)4233 backenable(queue_t *q, uchar_t pri)
4234 {
4235 	queue_t	*nq;
4236 
4237 	/*
4238 	 * Our presence might not prevent other modules in our own
4239 	 * stream from popping/pushing since the caller of getq might not
4240 	 * have a claim on the queue (some drivers do a getq on somebody
4241 	 * else's queue - they know that the queue itself is not going away
4242 	 * but the framework has to guarantee q_next in that stream).
4243 	 */
4244 	claimstr(q);
4245 
4246 	/* Find nearest back queue with service proc */
4247 	for (nq = backq(q); nq && !nq->q_qinfo->qi_srvp; nq = backq(nq)) {
4248 		ASSERT(STRMATED(q->q_stream) || STREAM(q) == STREAM(nq));
4249 	}
4250 
4251 	if (nq) {
4252 		kthread_t *freezer;
4253 		/*
4254 		 * backenable can be called either with no locks held
4255 		 * or with the stream frozen (the latter occurs when a module
4256 		 * calls rmvq with the stream frozen). If the stream is frozen
4257 		 * by the caller the caller will hold all qlocks in the stream.
4258 		 * Note that a frozen stream doesn't freeze a mated stream,
4259 		 * so we explicitly check for that.
4260 		 */
4261 		freezer = STREAM(q)->sd_freezer;
4262 		if (freezer != curthread || STREAM(q) != STREAM(nq)) {
4263 			mutex_enter(QLOCK(nq));
4264 		}
4265 #ifdef DEBUG
4266 		else {
4267 			ASSERT(frozenstr(q));
4268 			ASSERT(MUTEX_HELD(QLOCK(q)));
4269 			ASSERT(MUTEX_HELD(QLOCK(nq)));
4270 		}
4271 #endif
4272 		setqback(nq, pri);
4273 		qenable_locked(nq);
4274 		if (freezer != curthread || STREAM(q) != STREAM(nq))
4275 			mutex_exit(QLOCK(nq));
4276 	}
4277 	releasestr(q);
4278 }
4279 
4280 /*
4281  * Return the appropriate errno when one of flags_to_check is set
4282  * in sd_flags. Uses the exported error routines if they are set.
4283  * Will return 0 if non error is set (or if the exported error routines
4284  * do not return an error).
4285  *
4286  * If there is both a read and write error to check, we prefer the read error.
4287  * Also, give preference to recorded errno's over the error functions.
4288  * The flags that are handled are:
4289  *	STPLEX		return EINVAL
4290  *	STRDERR		return sd_rerror (and clear if STRDERRNONPERSIST)
4291  *	STWRERR		return sd_werror (and clear if STWRERRNONPERSIST)
4292  *	STRHUP		return sd_werror
4293  *
4294  * If the caller indicates that the operation is a peek, a nonpersistent error
4295  * is not cleared.
4296  */
4297 int
strgeterr(stdata_t * stp,int32_t flags_to_check,int ispeek)4298 strgeterr(stdata_t *stp, int32_t flags_to_check, int ispeek)
4299 {
4300 	int32_t sd_flag = stp->sd_flag & flags_to_check;
4301 	int error = 0;
4302 
4303 	ASSERT(MUTEX_HELD(&stp->sd_lock));
4304 	ASSERT((flags_to_check & ~(STRDERR|STWRERR|STRHUP|STPLEX)) == 0);
4305 	if (sd_flag & STPLEX)
4306 		error = EINVAL;
4307 	else if (sd_flag & STRDERR) {
4308 		error = stp->sd_rerror;
4309 		if ((stp->sd_flag & STRDERRNONPERSIST) && !ispeek) {
4310 			/*
4311 			 * Read errors are non-persistent i.e. discarded once
4312 			 * returned to a non-peeking caller,
4313 			 */
4314 			stp->sd_rerror = 0;
4315 			stp->sd_flag &= ~STRDERR;
4316 		}
4317 		if (error == 0 && stp->sd_rderrfunc != NULL) {
4318 			int clearerr = 0;
4319 
4320 			error = (*stp->sd_rderrfunc)(stp->sd_vnode, ispeek,
4321 			    &clearerr);
4322 			if (clearerr) {
4323 				stp->sd_flag &= ~STRDERR;
4324 				stp->sd_rderrfunc = NULL;
4325 			}
4326 		}
4327 	} else if (sd_flag & STWRERR) {
4328 		error = stp->sd_werror;
4329 		if ((stp->sd_flag & STWRERRNONPERSIST) && !ispeek) {
4330 			/*
4331 			 * Write errors are non-persistent i.e. discarded once
4332 			 * returned to a non-peeking caller,
4333 			 */
4334 			stp->sd_werror = 0;
4335 			stp->sd_flag &= ~STWRERR;
4336 		}
4337 		if (error == 0 && stp->sd_wrerrfunc != NULL) {
4338 			int clearerr = 0;
4339 
4340 			error = (*stp->sd_wrerrfunc)(stp->sd_vnode, ispeek,
4341 			    &clearerr);
4342 			if (clearerr) {
4343 				stp->sd_flag &= ~STWRERR;
4344 				stp->sd_wrerrfunc = NULL;
4345 			}
4346 		}
4347 	} else if (sd_flag & STRHUP) {
4348 		/* sd_werror set when STRHUP */
4349 		error = stp->sd_werror;
4350 	}
4351 	return (error);
4352 }
4353 
4354 
4355 /*
4356  * Single-thread open/close/push/pop
4357  * for twisted streams also
4358  */
4359 int
strstartplumb(stdata_t * stp,int flag,int cmd)4360 strstartplumb(stdata_t *stp, int flag, int cmd)
4361 {
4362 	int waited = 1;
4363 	int error = 0;
4364 
4365 	if (STRMATED(stp)) {
4366 		struct stdata *stmatep = stp->sd_mate;
4367 
4368 		STRLOCKMATES(stp);
4369 		while (waited) {
4370 			waited = 0;
4371 			while (stmatep->sd_flag & (STWOPEN|STRCLOSE|STRPLUMB)) {
4372 				if ((cmd == I_POP) &&
4373 				    (flag & (FNDELAY|FNONBLOCK))) {
4374 					STRUNLOCKMATES(stp);
4375 					return (EAGAIN);
4376 				}
4377 				waited = 1;
4378 				mutex_exit(&stp->sd_lock);
4379 				if (!cv_wait_sig(&stmatep->sd_monitor,
4380 				    &stmatep->sd_lock)) {
4381 					mutex_exit(&stmatep->sd_lock);
4382 					return (EINTR);
4383 				}
4384 				mutex_exit(&stmatep->sd_lock);
4385 				STRLOCKMATES(stp);
4386 			}
4387 			while (stp->sd_flag & (STWOPEN|STRCLOSE|STRPLUMB)) {
4388 				if ((cmd == I_POP) &&
4389 				    (flag & (FNDELAY|FNONBLOCK))) {
4390 					STRUNLOCKMATES(stp);
4391 					return (EAGAIN);
4392 				}
4393 				waited = 1;
4394 				mutex_exit(&stmatep->sd_lock);
4395 				if (!cv_wait_sig(&stp->sd_monitor,
4396 				    &stp->sd_lock)) {
4397 					mutex_exit(&stp->sd_lock);
4398 					return (EINTR);
4399 				}
4400 				mutex_exit(&stp->sd_lock);
4401 				STRLOCKMATES(stp);
4402 			}
4403 			if (stp->sd_flag & (STRDERR|STWRERR|STRHUP|STPLEX)) {
4404 				error = strgeterr(stp,
4405 				    STRDERR|STWRERR|STRHUP|STPLEX, 0);
4406 				if (error != 0) {
4407 					STRUNLOCKMATES(stp);
4408 					return (error);
4409 				}
4410 			}
4411 		}
4412 		stp->sd_flag |= STRPLUMB;
4413 		STRUNLOCKMATES(stp);
4414 	} else {
4415 		mutex_enter(&stp->sd_lock);
4416 		while (stp->sd_flag & (STWOPEN|STRCLOSE|STRPLUMB)) {
4417 			if (((cmd == I_POP) || (cmd == _I_REMOVE)) &&
4418 			    (flag & (FNDELAY|FNONBLOCK))) {
4419 				mutex_exit(&stp->sd_lock);
4420 				return (EAGAIN);
4421 			}
4422 			if (!cv_wait_sig(&stp->sd_monitor, &stp->sd_lock)) {
4423 				mutex_exit(&stp->sd_lock);
4424 				return (EINTR);
4425 			}
4426 			if (stp->sd_flag & (STRDERR|STWRERR|STRHUP|STPLEX)) {
4427 				error = strgeterr(stp,
4428 				    STRDERR|STWRERR|STRHUP|STPLEX, 0);
4429 				if (error != 0) {
4430 					mutex_exit(&stp->sd_lock);
4431 					return (error);
4432 				}
4433 			}
4434 		}
4435 		stp->sd_flag |= STRPLUMB;
4436 		mutex_exit(&stp->sd_lock);
4437 	}
4438 	return (0);
4439 }
4440 
4441 /*
4442  * Complete the plumbing operation associated with stream `stp'.
4443  */
4444 void
strendplumb(stdata_t * stp)4445 strendplumb(stdata_t *stp)
4446 {
4447 	ASSERT(MUTEX_HELD(&stp->sd_lock));
4448 	ASSERT(stp->sd_flag & STRPLUMB);
4449 	stp->sd_flag &= ~STRPLUMB;
4450 	cv_broadcast(&stp->sd_monitor);
4451 }
4452 
4453 /*
4454  * This describes how the STREAMS framework handles synchronization
4455  * during open/push and close/pop.
4456  * The key interfaces for open and close are qprocson and qprocsoff,
4457  * respectively. While the close case in general is harder both open
4458  * have close have significant similarities.
4459  *
4460  * During close the STREAMS framework has to both ensure that there
4461  * are no stale references to the queue pair (and syncq) that
4462  * are being closed and also provide the guarantees that are documented
4463  * in qprocsoff(9F).
4464  * If there are stale references to the queue that is closing it can
4465  * result in kernel memory corruption or kernel panics.
4466  *
4467  * Note that is it up to the module/driver to ensure that it itself
4468  * does not have any stale references to the closing queues once its close
4469  * routine returns. This includes:
4470  *  - Cancelling any timeout/bufcall/qtimeout/qbufcall callback routines
4471  *    associated with the queues. For timeout and bufcall callbacks the
4472  *    module/driver also has to ensure (or wait for) any callbacks that
4473  *    are in progress.
4474  *  - If the module/driver is using esballoc it has to ensure that any
4475  *    esballoc free functions do not refer to a queue that has closed.
4476  *    (Note that in general the close routine can not wait for the esballoc'ed
4477  *    messages to be freed since that can cause a deadlock.)
4478  *  - Cancelling any interrupts that refer to the closing queues and
4479  *    also ensuring that there are no interrupts in progress that will
4480  *    refer to the closing queues once the close routine returns.
4481  *  - For multiplexors removing any driver global state that refers to
4482  *    the closing queue and also ensuring that there are no threads in
4483  *    the multiplexor that has picked up a queue pointer but not yet
4484  *    finished using it.
4485  *
4486  * In addition, a driver/module can only reference the q_next pointer
4487  * in its open, close, put, or service procedures or in a
4488  * qtimeout/qbufcall callback procedure executing "on" the correct
4489  * stream. Thus it can not reference the q_next pointer in an interrupt
4490  * routine or a timeout, bufcall or esballoc callback routine. Likewise
4491  * it can not reference q_next of a different queue e.g. in a mux that
4492  * passes messages from one queues put/service procedure to another queue.
4493  * In all the cases when the driver/module can not access the q_next
4494  * field it must use the *next* versions e.g. canputnext instead of
4495  * canput(q->q_next) and putnextctl instead of putctl(q->q_next, ...).
4496  *
4497  *
4498  * Assuming that the driver/module conforms to the above constraints
4499  * the STREAMS framework has to avoid stale references to q_next for all
4500  * the framework internal cases which include (but are not limited to):
4501  *  - Threads in canput/canputnext/backenable and elsewhere that are
4502  *    walking q_next.
4503  *  - Messages on a syncq that have a reference to the queue through b_queue.
4504  *  - Messages on an outer perimeter (syncq) that have a reference to the
4505  *    queue through b_queue.
4506  *  - Threads that use q_nfsrv (e.g. canput) to find a queue.
4507  *    Note that only canput and bcanput use q_nfsrv without any locking.
4508  *
4509  * The STREAMS framework providing the qprocsoff(9F) guarantees means that
4510  * after qprocsoff returns, the framework has to ensure that no threads can
4511  * enter the put or service routines for the closing read or write-side queue.
4512  * In addition to preventing "direct" entry into the put procedures
4513  * the framework also has to prevent messages being drained from
4514  * the syncq or the outer perimeter.
4515  * XXX Note that currently qdetach does relies on D_MTOCEXCL as the only
4516  * mechanism to prevent qwriter(PERIM_OUTER) from running after
4517  * qprocsoff has returned.
4518  * Note that if a module/driver uses put(9F) on one of its own queues
4519  * it is up to the module/driver to ensure that the put() doesn't
4520  * get called when the queue is closing.
4521  *
4522  *
4523  * The framework aspects of the above "contract" is implemented by
4524  * qprocsoff, removeq, and strlock:
4525  *  - qprocsoff (disable_svc) sets QWCLOSE to prevent runservice from
4526  *    entering the service procedures.
4527  *  - strlock acquires the sd_lock and sd_reflock to prevent putnext,
4528  *    canputnext, backenable etc from dereferencing the q_next that will
4529  *    soon change.
4530  *  - strlock waits for sd_refcnt to be zero to wait for e.g. any canputnext
4531  *    or other q_next walker that uses claimstr/releasestr to finish.
4532  *  - optionally for every syncq in the stream strlock acquires all the
4533  *    sq_lock's and waits for all sq_counts to drop to a value that indicates
4534  *    that no thread executes in the put or service procedures and that no
4535  *    thread is draining into the module/driver. This ensures that no
4536  *    open, close, put, service, or qtimeout/qbufcall callback procedure is
4537  *    currently executing hence no such thread can end up with the old stale
4538  *    q_next value and no canput/backenable can have the old stale
4539  *    q_nfsrv/q_next.
4540  *  - qdetach (wait_svc) makes sure that any scheduled or running threads
4541  *    have either finished or observed the QWCLOSE flag and gone away.
4542  */
4543 
4544 
4545 /*
4546  * Get all the locks necessary to change q_next.
4547  *
4548  * Wait for sd_refcnt to reach 0 and, if sqlist is present, wait for the
4549  * sq_count of each syncq in the list to drop to sq_rmqcount, indicating that
4550  * the only threads inside the syncq are threads currently calling removeq().
4551  * Since threads calling removeq() are in the process of removing their queues
4552  * from the stream, we do not need to worry about them accessing a stale q_next
4553  * pointer and thus we do not need to wait for them to exit (in fact, waiting
4554  * for them can cause deadlock).
4555  *
4556  * This routine is subject to starvation since it does not set any flag to
4557  * prevent threads from entering a module in the stream (i.e. sq_count can
4558  * increase on some syncq while it is waiting on some other syncq).
4559  *
4560  * Assumes that only one thread attempts to call strlock for a given
4561  * stream. If this is not the case the two threads would deadlock.
4562  * This assumption is guaranteed since strlock is only called by insertq
4563  * and removeq and streams plumbing changes are single-threaded for
4564  * a given stream using the STWOPEN, STRCLOSE, and STRPLUMB flags.
4565  *
4566  * For pipes, it is not difficult to atomically designate a pair of streams
4567  * to be mated. Once mated atomically by the framework the twisted pair remain
4568  * configured that way until dismantled atomically by the framework.
4569  * When plumbing takes place on a twisted stream it is necessary to ensure that
4570  * this operation is done exclusively on the twisted stream since two such
4571  * operations, each initiated on different ends of the pipe will deadlock
4572  * waiting for each other to complete.
4573  *
4574  * On entry, no locks should be held.
4575  * The locks acquired and held by strlock depends on a few factors.
4576  * - If sqlist is non-NULL all the syncq locks in the sqlist will be acquired
4577  *   and held on exit and all sq_count are at an acceptable level.
4578  * - In all cases, sd_lock and sd_reflock are acquired and held on exit with
4579  *   sd_refcnt being zero.
4580  */
4581 
4582 static void
strlock(struct stdata * stp,sqlist_t * sqlist)4583 strlock(struct stdata *stp, sqlist_t *sqlist)
4584 {
4585 	syncql_t *sql, *sql2;
4586 retry:
4587 	/*
4588 	 * Wait for any claimstr to go away.
4589 	 */
4590 	if (STRMATED(stp)) {
4591 		struct stdata *stp1, *stp2;
4592 
4593 		STRLOCKMATES(stp);
4594 		/*
4595 		 * Note that the selection of locking order is not
4596 		 * important, just that they are always acquired in
4597 		 * the same order.  To assure this, we choose this
4598 		 * order based on the value of the pointer, and since
4599 		 * the pointer will not change for the life of this
4600 		 * pair, we will always grab the locks in the same
4601 		 * order (and hence, prevent deadlocks).
4602 		 */
4603 		if (&(stp->sd_lock) > &((stp->sd_mate)->sd_lock)) {
4604 			stp1 = stp;
4605 			stp2 = stp->sd_mate;
4606 		} else {
4607 			stp2 = stp;
4608 			stp1 = stp->sd_mate;
4609 		}
4610 		mutex_enter(&stp1->sd_reflock);
4611 		if (stp1->sd_refcnt > 0) {
4612 			STRUNLOCKMATES(stp);
4613 			cv_wait(&stp1->sd_refmonitor, &stp1->sd_reflock);
4614 			mutex_exit(&stp1->sd_reflock);
4615 			goto retry;
4616 		}
4617 		mutex_enter(&stp2->sd_reflock);
4618 		if (stp2->sd_refcnt > 0) {
4619 			STRUNLOCKMATES(stp);
4620 			mutex_exit(&stp1->sd_reflock);
4621 			cv_wait(&stp2->sd_refmonitor, &stp2->sd_reflock);
4622 			mutex_exit(&stp2->sd_reflock);
4623 			goto retry;
4624 		}
4625 		STREAM_PUTLOCKS_ENTER(stp1);
4626 		STREAM_PUTLOCKS_ENTER(stp2);
4627 	} else {
4628 		mutex_enter(&stp->sd_lock);
4629 		mutex_enter(&stp->sd_reflock);
4630 		while (stp->sd_refcnt > 0) {
4631 			mutex_exit(&stp->sd_lock);
4632 			cv_wait(&stp->sd_refmonitor, &stp->sd_reflock);
4633 			if (mutex_tryenter(&stp->sd_lock) == 0) {
4634 				mutex_exit(&stp->sd_reflock);
4635 				mutex_enter(&stp->sd_lock);
4636 				mutex_enter(&stp->sd_reflock);
4637 			}
4638 		}
4639 		STREAM_PUTLOCKS_ENTER(stp);
4640 	}
4641 
4642 	if (sqlist == NULL)
4643 		return;
4644 
4645 	for (sql = sqlist->sqlist_head; sql; sql = sql->sql_next) {
4646 		syncq_t *sq = sql->sql_sq;
4647 		uint16_t count;
4648 
4649 		mutex_enter(SQLOCK(sq));
4650 		count = sq->sq_count;
4651 		ASSERT(sq->sq_rmqcount <= count);
4652 		SQ_PUTLOCKS_ENTER(sq);
4653 		SUM_SQ_PUTCOUNTS(sq, count);
4654 		if (count == sq->sq_rmqcount)
4655 			continue;
4656 
4657 		/* Failed - drop all locks that we have acquired so far */
4658 		if (STRMATED(stp)) {
4659 			STREAM_PUTLOCKS_EXIT(stp);
4660 			STREAM_PUTLOCKS_EXIT(stp->sd_mate);
4661 			STRUNLOCKMATES(stp);
4662 			mutex_exit(&stp->sd_reflock);
4663 			mutex_exit(&stp->sd_mate->sd_reflock);
4664 		} else {
4665 			STREAM_PUTLOCKS_EXIT(stp);
4666 			mutex_exit(&stp->sd_lock);
4667 			mutex_exit(&stp->sd_reflock);
4668 		}
4669 		for (sql2 = sqlist->sqlist_head; sql2 != sql;
4670 		    sql2 = sql2->sql_next) {
4671 			SQ_PUTLOCKS_EXIT(sql2->sql_sq);
4672 			mutex_exit(SQLOCK(sql2->sql_sq));
4673 		}
4674 
4675 		/*
4676 		 * The wait loop below may starve when there are many threads
4677 		 * claiming the syncq. This is especially a problem with permod
4678 		 * syncqs (IP). To lessen the impact of the problem we increment
4679 		 * sq_needexcl and clear fastbits so that putnexts will slow
4680 		 * down and call sqenable instead of draining right away.
4681 		 */
4682 		sq->sq_needexcl++;
4683 		SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
4684 		while (count > sq->sq_rmqcount) {
4685 			sq->sq_flags |= SQ_WANTWAKEUP;
4686 			SQ_PUTLOCKS_EXIT(sq);
4687 			cv_wait(&sq->sq_wait, SQLOCK(sq));
4688 			count = sq->sq_count;
4689 			SQ_PUTLOCKS_ENTER(sq);
4690 			SUM_SQ_PUTCOUNTS(sq, count);
4691 		}
4692 		sq->sq_needexcl--;
4693 		if (sq->sq_needexcl == 0)
4694 			SQ_PUTCOUNT_SETFAST_LOCKED(sq);
4695 		SQ_PUTLOCKS_EXIT(sq);
4696 		ASSERT(count == sq->sq_rmqcount);
4697 		mutex_exit(SQLOCK(sq));
4698 		goto retry;
4699 	}
4700 }
4701 
4702 /*
4703  * Drop all the locks that strlock acquired.
4704  */
4705 static void
strunlock(struct stdata * stp,sqlist_t * sqlist)4706 strunlock(struct stdata *stp, sqlist_t *sqlist)
4707 {
4708 	syncql_t *sql;
4709 
4710 	if (STRMATED(stp)) {
4711 		STREAM_PUTLOCKS_EXIT(stp);
4712 		STREAM_PUTLOCKS_EXIT(stp->sd_mate);
4713 		STRUNLOCKMATES(stp);
4714 		mutex_exit(&stp->sd_reflock);
4715 		mutex_exit(&stp->sd_mate->sd_reflock);
4716 	} else {
4717 		STREAM_PUTLOCKS_EXIT(stp);
4718 		mutex_exit(&stp->sd_lock);
4719 		mutex_exit(&stp->sd_reflock);
4720 	}
4721 
4722 	if (sqlist == NULL)
4723 		return;
4724 
4725 	for (sql = sqlist->sqlist_head; sql; sql = sql->sql_next) {
4726 		SQ_PUTLOCKS_EXIT(sql->sql_sq);
4727 		mutex_exit(SQLOCK(sql->sql_sq));
4728 	}
4729 }
4730 
4731 /*
4732  * When the module has service procedure, we need check if the next
4733  * module which has service procedure is in flow control to trigger
4734  * the backenable.
4735  */
4736 static void
backenable_insertedq(queue_t * q)4737 backenable_insertedq(queue_t *q)
4738 {
4739 	qband_t	*qbp;
4740 
4741 	claimstr(q);
4742 	if (q->q_qinfo->qi_srvp != NULL && q->q_next != NULL) {
4743 		if (q->q_next->q_nfsrv->q_flag & QWANTW)
4744 			backenable(q, 0);
4745 
4746 		qbp = q->q_next->q_nfsrv->q_bandp;
4747 		for (; qbp != NULL; qbp = qbp->qb_next)
4748 			if ((qbp->qb_flag & QB_WANTW) && qbp->qb_first != NULL)
4749 				backenable(q, qbp->qb_first->b_band);
4750 	}
4751 	releasestr(q);
4752 }
4753 
4754 /*
4755  * Given two read queues, insert a new single one after another.
4756  *
4757  * This routine acquires all the necessary locks in order to change
4758  * q_next and related pointer using strlock().
4759  * It depends on the stream head ensuring that there are no concurrent
4760  * insertq or removeq on the same stream. The stream head ensures this
4761  * using the flags STWOPEN, STRCLOSE, and STRPLUMB.
4762  *
4763  * Note that no syncq locks are held during the q_next change. This is
4764  * applied to all streams since, unlike removeq, there is no problem of stale
4765  * pointers when adding a module to the stream. Thus drivers/modules that do a
4766  * canput(rq->q_next) would never get a closed/freed queue pointer even if we
4767  * applied this optimization to all streams.
4768  */
4769 void
insertq(struct stdata * stp,queue_t * new)4770 insertq(struct stdata *stp, queue_t *new)
4771 {
4772 	queue_t	*after;
4773 	queue_t *wafter;
4774 	queue_t *wnew = _WR(new);
4775 	boolean_t have_fifo = B_FALSE;
4776 
4777 	if (new->q_flag & _QINSERTING) {
4778 		ASSERT(stp->sd_vnode->v_type != VFIFO);
4779 		after = new->q_next;
4780 		wafter = _WR(new->q_next);
4781 	} else {
4782 		after = _RD(stp->sd_wrq);
4783 		wafter = stp->sd_wrq;
4784 	}
4785 
4786 	TRACE_2(TR_FAC_STREAMS_FR, TR_INSERTQ,
4787 	    "insertq:%p, %p", after, new);
4788 	ASSERT(after->q_flag & QREADR);
4789 	ASSERT(new->q_flag & QREADR);
4790 
4791 	strlock(stp, NULL);
4792 
4793 	/* Do we have a FIFO? */
4794 	if (wafter->q_next == after) {
4795 		have_fifo = B_TRUE;
4796 		wnew->q_next = new;
4797 	} else {
4798 		wnew->q_next = wafter->q_next;
4799 	}
4800 	new->q_next = after;
4801 
4802 	set_nfsrv_ptr(new, wnew, after, wafter);
4803 	/*
4804 	 * set_nfsrv_ptr() needs to know if this is an insertion or not,
4805 	 * so only reset this flag after calling it.
4806 	 */
4807 	new->q_flag &= ~_QINSERTING;
4808 
4809 	if (have_fifo) {
4810 		wafter->q_next = wnew;
4811 	} else {
4812 		if (wafter->q_next)
4813 			_OTHERQ(wafter->q_next)->q_next = new;
4814 		wafter->q_next = wnew;
4815 	}
4816 
4817 	set_qend(new);
4818 	/* The QEND flag might have to be updated for the upstream guy */
4819 	set_qend(after);
4820 
4821 	ASSERT(_SAMESTR(new) == O_SAMESTR(new));
4822 	ASSERT(_SAMESTR(wnew) == O_SAMESTR(wnew));
4823 	ASSERT(_SAMESTR(after) == O_SAMESTR(after));
4824 	ASSERT(_SAMESTR(wafter) == O_SAMESTR(wafter));
4825 	strsetuio(stp);
4826 
4827 	/*
4828 	 * If this was a module insertion, bump the push count.
4829 	 */
4830 	if (!(new->q_flag & QISDRV))
4831 		stp->sd_pushcnt++;
4832 
4833 	strunlock(stp, NULL);
4834 
4835 	/* check if the write Q needs backenable */
4836 	backenable_insertedq(wnew);
4837 
4838 	/* check if the read Q needs backenable */
4839 	backenable_insertedq(new);
4840 }
4841 
4842 /*
4843  * Given a read queue, unlink it from any neighbors.
4844  *
4845  * This routine acquires all the necessary locks in order to
4846  * change q_next and related pointers and also guard against
4847  * stale references (e.g. through q_next) to the queue that
4848  * is being removed. It also plays part of the role in ensuring
4849  * that the module's/driver's put procedure doesn't get called
4850  * after qprocsoff returns.
4851  *
4852  * Removeq depends on the stream head ensuring that there are
4853  * no concurrent insertq or removeq on the same stream. The
4854  * stream head ensures this using the flags STWOPEN, STRCLOSE and
4855  * STRPLUMB.
4856  *
4857  * The set of locks needed to remove the queue is different in
4858  * different cases:
4859  *
4860  * Acquire sd_lock, sd_reflock, and all the syncq locks in the stream after
4861  * waiting for the syncq reference count to drop to 0 indicating that no
4862  * non-close threads are present anywhere in the stream. This ensures that any
4863  * module/driver can reference q_next in its open, close, put, or service
4864  * procedures.
4865  *
4866  * The sq_rmqcount counter tracks the number of threads inside removeq().
4867  * strlock() ensures that there is either no threads executing inside perimeter
4868  * or there is only a thread calling qprocsoff().
4869  *
4870  * strlock() compares the value of sq_count with the number of threads inside
4871  * removeq() and waits until sq_count is equal to sq_rmqcount. We need to wakeup
4872  * any threads waiting in strlock() when the sq_rmqcount increases.
4873  */
4874 
4875 void
removeq(queue_t * qp)4876 removeq(queue_t *qp)
4877 {
4878 	queue_t *wqp = _WR(qp);
4879 	struct stdata *stp = STREAM(qp);
4880 	sqlist_t *sqlist = NULL;
4881 	boolean_t isdriver;
4882 	int moved;
4883 	syncq_t *sq = qp->q_syncq;
4884 	syncq_t *wsq = wqp->q_syncq;
4885 
4886 	ASSERT(stp);
4887 
4888 	TRACE_2(TR_FAC_STREAMS_FR, TR_REMOVEQ,
4889 	    "removeq:%p %p", qp, wqp);
4890 	ASSERT(qp->q_flag&QREADR);
4891 
4892 	/*
4893 	 * For queues using Synchronous streams, we must wait for all threads in
4894 	 * rwnext() to drain out before proceeding.
4895 	 */
4896 	if (qp->q_flag & QSYNCSTR) {
4897 		/* First, we need wakeup any threads blocked in rwnext() */
4898 		mutex_enter(SQLOCK(sq));
4899 		if (sq->sq_flags & SQ_WANTWAKEUP) {
4900 			sq->sq_flags &= ~SQ_WANTWAKEUP;
4901 			cv_broadcast(&sq->sq_wait);
4902 		}
4903 		mutex_exit(SQLOCK(sq));
4904 
4905 		if (wsq != sq) {
4906 			mutex_enter(SQLOCK(wsq));
4907 			if (wsq->sq_flags & SQ_WANTWAKEUP) {
4908 				wsq->sq_flags &= ~SQ_WANTWAKEUP;
4909 				cv_broadcast(&wsq->sq_wait);
4910 			}
4911 			mutex_exit(SQLOCK(wsq));
4912 		}
4913 
4914 		mutex_enter(QLOCK(qp));
4915 		while (qp->q_rwcnt > 0) {
4916 			qp->q_flag |= QWANTRMQSYNC;
4917 			cv_wait(&qp->q_wait, QLOCK(qp));
4918 		}
4919 		mutex_exit(QLOCK(qp));
4920 
4921 		mutex_enter(QLOCK(wqp));
4922 		while (wqp->q_rwcnt > 0) {
4923 			wqp->q_flag |= QWANTRMQSYNC;
4924 			cv_wait(&wqp->q_wait, QLOCK(wqp));
4925 		}
4926 		mutex_exit(QLOCK(wqp));
4927 	}
4928 
4929 	mutex_enter(SQLOCK(sq));
4930 	sq->sq_rmqcount++;
4931 	if (sq->sq_flags & SQ_WANTWAKEUP) {
4932 		sq->sq_flags &= ~SQ_WANTWAKEUP;
4933 		cv_broadcast(&sq->sq_wait);
4934 	}
4935 	mutex_exit(SQLOCK(sq));
4936 
4937 	isdriver = (qp->q_flag & QISDRV);
4938 
4939 	sqlist = sqlist_build(qp, stp, STRMATED(stp));
4940 	strlock(stp, sqlist);
4941 
4942 	reset_nfsrv_ptr(qp, wqp);
4943 
4944 	ASSERT(wqp->q_next == NULL || backq(qp)->q_next == qp);
4945 	ASSERT(qp->q_next == NULL || backq(wqp)->q_next == wqp);
4946 	/* Do we have a FIFO? */
4947 	if (wqp->q_next == qp) {
4948 		stp->sd_wrq->q_next = _RD(stp->sd_wrq);
4949 	} else {
4950 		if (wqp->q_next)
4951 			backq(qp)->q_next = qp->q_next;
4952 		if (qp->q_next)
4953 			backq(wqp)->q_next = wqp->q_next;
4954 	}
4955 
4956 	/* The QEND flag might have to be updated for the upstream guy */
4957 	if (qp->q_next)
4958 		set_qend(qp->q_next);
4959 
4960 	ASSERT(_SAMESTR(stp->sd_wrq) == O_SAMESTR(stp->sd_wrq));
4961 	ASSERT(_SAMESTR(_RD(stp->sd_wrq)) == O_SAMESTR(_RD(stp->sd_wrq)));
4962 
4963 	/*
4964 	 * Move any messages destined for the put procedures to the next
4965 	 * syncq in line. Otherwise free them.
4966 	 */
4967 	moved = 0;
4968 	/*
4969 	 * Quick check to see whether there are any messages or events.
4970 	 */
4971 	if (qp->q_syncqmsgs != 0 || (qp->q_syncq->sq_flags & SQ_EVENTS))
4972 		moved += propagate_syncq(qp);
4973 	if (wqp->q_syncqmsgs != 0 ||
4974 	    (wqp->q_syncq->sq_flags & SQ_EVENTS))
4975 		moved += propagate_syncq(wqp);
4976 
4977 	strsetuio(stp);
4978 
4979 	/*
4980 	 * If this was a module removal, decrement the push count.
4981 	 */
4982 	if (!isdriver)
4983 		stp->sd_pushcnt--;
4984 
4985 	strunlock(stp, sqlist);
4986 	sqlist_free(sqlist);
4987 
4988 	/*
4989 	 * Make sure any messages that were propagated are drained.
4990 	 * Also clear any QFULL bit caused by messages that were propagated.
4991 	 */
4992 
4993 	if (qp->q_next != NULL) {
4994 		clr_qfull(qp);
4995 		/*
4996 		 * For the driver calling qprocsoff, propagate_syncq
4997 		 * frees all the messages instead of putting it in
4998 		 * the stream head
4999 		 */
5000 		if (!isdriver && (moved > 0))
5001 			emptysq(qp->q_next->q_syncq);
5002 	}
5003 	if (wqp->q_next != NULL) {
5004 		clr_qfull(wqp);
5005 		/*
5006 		 * We come here for any pop of a module except for the
5007 		 * case of driver being removed. We don't call emptysq
5008 		 * if we did not move any messages. This will avoid holding
5009 		 * PERMOD syncq locks in emptysq
5010 		 */
5011 		if (moved > 0)
5012 			emptysq(wqp->q_next->q_syncq);
5013 	}
5014 
5015 	mutex_enter(SQLOCK(sq));
5016 	sq->sq_rmqcount--;
5017 	mutex_exit(SQLOCK(sq));
5018 }
5019 
5020 /*
5021  * Prevent further entry by setting a flag (like SQ_FROZEN, SQ_BLOCKED or
5022  * SQ_WRITER) on a syncq.
5023  * If maxcnt is not -1 it assumes that caller has "maxcnt" claim(s) on the
5024  * sync queue and waits until sq_count reaches maxcnt.
5025  *
5026  * If maxcnt is -1 there's no need to grab sq_putlocks since the caller
5027  * does not care about putnext threads that are in the middle of calling put
5028  * entry points.
5029  *
5030  * This routine is used for both inner and outer syncqs.
5031  */
5032 static void
blocksq(syncq_t * sq,ushort_t flag,int maxcnt)5033 blocksq(syncq_t *sq, ushort_t flag, int maxcnt)
5034 {
5035 	uint16_t count = 0;
5036 
5037 	mutex_enter(SQLOCK(sq));
5038 	/*
5039 	 * Wait for SQ_FROZEN/SQ_BLOCKED to be reset.
5040 	 * SQ_FROZEN will be set if there is a frozen stream that has a
5041 	 * queue which also refers to this "shared" syncq.
5042 	 * SQ_BLOCKED will be set if there is "off" queue which also
5043 	 * refers to this "shared" syncq.
5044 	 */
5045 	if (maxcnt != -1) {
5046 		count = sq->sq_count;
5047 		SQ_PUTLOCKS_ENTER(sq);
5048 		SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
5049 		SUM_SQ_PUTCOUNTS(sq, count);
5050 	}
5051 	sq->sq_needexcl++;
5052 	ASSERT(sq->sq_needexcl != 0);	/* wraparound */
5053 
5054 	while ((sq->sq_flags & flag) ||
5055 	    (maxcnt != -1 && count > (unsigned)maxcnt)) {
5056 		sq->sq_flags |= SQ_WANTWAKEUP;
5057 		if (maxcnt != -1) {
5058 			SQ_PUTLOCKS_EXIT(sq);
5059 		}
5060 		cv_wait(&sq->sq_wait, SQLOCK(sq));
5061 		if (maxcnt != -1) {
5062 			count = sq->sq_count;
5063 			SQ_PUTLOCKS_ENTER(sq);
5064 			SUM_SQ_PUTCOUNTS(sq, count);
5065 		}
5066 	}
5067 	sq->sq_needexcl--;
5068 	sq->sq_flags |= flag;
5069 	ASSERT(maxcnt == -1 || count == maxcnt);
5070 	if (maxcnt != -1) {
5071 		if (sq->sq_needexcl == 0) {
5072 			SQ_PUTCOUNT_SETFAST_LOCKED(sq);
5073 		}
5074 		SQ_PUTLOCKS_EXIT(sq);
5075 	} else if (sq->sq_needexcl == 0) {
5076 		SQ_PUTCOUNT_SETFAST(sq);
5077 	}
5078 
5079 	mutex_exit(SQLOCK(sq));
5080 }
5081 
5082 /*
5083  * Reset a flag that was set with blocksq.
5084  *
5085  * Can not use this routine to reset SQ_WRITER.
5086  *
5087  * If "isouter" is set then the syncq is assumed to be an outer perimeter
5088  * and drain_syncq is not called. Instead we rely on the qwriter_outer thread
5089  * to handle the queued qwriter operations.
5090  *
5091  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5092  * sq_putlocks are used.
5093  */
5094 static void
unblocksq(syncq_t * sq,uint16_t resetflag,int isouter)5095 unblocksq(syncq_t *sq, uint16_t resetflag, int isouter)
5096 {
5097 	uint16_t flags;
5098 
5099 	mutex_enter(SQLOCK(sq));
5100 	ASSERT(resetflag != SQ_WRITER);
5101 	ASSERT(sq->sq_flags & resetflag);
5102 	flags = sq->sq_flags & ~resetflag;
5103 	sq->sq_flags = flags;
5104 	if (flags & (SQ_QUEUED | SQ_WANTWAKEUP)) {
5105 		if (flags & SQ_WANTWAKEUP) {
5106 			flags &= ~SQ_WANTWAKEUP;
5107 			cv_broadcast(&sq->sq_wait);
5108 		}
5109 		sq->sq_flags = flags;
5110 		if ((flags & SQ_QUEUED) && !(flags & (SQ_STAYAWAY|SQ_EXCL))) {
5111 			if (!isouter) {
5112 				/* drain_syncq drops SQLOCK */
5113 				drain_syncq(sq);
5114 				return;
5115 			}
5116 		}
5117 	}
5118 	mutex_exit(SQLOCK(sq));
5119 }
5120 
5121 /*
5122  * Reset a flag that was set with blocksq.
5123  * Does not drain the syncq. Use emptysq() for that.
5124  * Returns 1 if SQ_QUEUED is set. Otherwise 0.
5125  *
5126  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5127  * sq_putlocks are used.
5128  */
5129 static int
dropsq(syncq_t * sq,uint16_t resetflag)5130 dropsq(syncq_t *sq, uint16_t resetflag)
5131 {
5132 	uint16_t flags;
5133 
5134 	mutex_enter(SQLOCK(sq));
5135 	ASSERT(sq->sq_flags & resetflag);
5136 	flags = sq->sq_flags & ~resetflag;
5137 	if (flags & SQ_WANTWAKEUP) {
5138 		flags &= ~SQ_WANTWAKEUP;
5139 		cv_broadcast(&sq->sq_wait);
5140 	}
5141 	sq->sq_flags = flags;
5142 	mutex_exit(SQLOCK(sq));
5143 	if (flags & SQ_QUEUED)
5144 		return (1);
5145 	return (0);
5146 }
5147 
5148 /*
5149  * Empty all the messages on a syncq.
5150  *
5151  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5152  * sq_putlocks are used.
5153  */
5154 static void
emptysq(syncq_t * sq)5155 emptysq(syncq_t *sq)
5156 {
5157 	uint16_t flags;
5158 
5159 	mutex_enter(SQLOCK(sq));
5160 	flags = sq->sq_flags;
5161 	if ((flags & SQ_QUEUED) && !(flags & (SQ_STAYAWAY|SQ_EXCL))) {
5162 		/*
5163 		 * To prevent potential recursive invocation of drain_syncq we
5164 		 * do not call drain_syncq if count is non-zero.
5165 		 */
5166 		if (sq->sq_count == 0) {
5167 			/* drain_syncq() drops SQLOCK */
5168 			drain_syncq(sq);
5169 			return;
5170 		} else
5171 			sqenable(sq);
5172 	}
5173 	mutex_exit(SQLOCK(sq));
5174 }
5175 
5176 /*
5177  * Ordered insert while removing duplicates.
5178  */
5179 static void
sqlist_insert(sqlist_t * sqlist,syncq_t * sqp)5180 sqlist_insert(sqlist_t *sqlist, syncq_t *sqp)
5181 {
5182 	syncql_t *sqlp, **prev_sqlpp, *new_sqlp;
5183 
5184 	prev_sqlpp = &sqlist->sqlist_head;
5185 	while ((sqlp = *prev_sqlpp) != NULL) {
5186 		if (sqlp->sql_sq >= sqp) {
5187 			if (sqlp->sql_sq == sqp)	/* duplicate */
5188 				return;
5189 			break;
5190 		}
5191 		prev_sqlpp = &sqlp->sql_next;
5192 	}
5193 	new_sqlp = &sqlist->sqlist_array[sqlist->sqlist_index++];
5194 	ASSERT((char *)new_sqlp < (char *)sqlist + sqlist->sqlist_size);
5195 	new_sqlp->sql_next = sqlp;
5196 	new_sqlp->sql_sq = sqp;
5197 	*prev_sqlpp = new_sqlp;
5198 }
5199 
5200 /*
5201  * Walk the write side queues until we hit either the driver
5202  * or a twist in the stream (_SAMESTR will return false in both
5203  * these cases) then turn around and walk the read side queues
5204  * back up to the stream head.
5205  */
5206 static void
sqlist_insertall(sqlist_t * sqlist,queue_t * q)5207 sqlist_insertall(sqlist_t *sqlist, queue_t *q)
5208 {
5209 	while (q != NULL) {
5210 		sqlist_insert(sqlist, q->q_syncq);
5211 
5212 		if (_SAMESTR(q))
5213 			q = q->q_next;
5214 		else if (!(q->q_flag & QREADR))
5215 			q = _RD(q);
5216 		else
5217 			q = NULL;
5218 	}
5219 }
5220 
5221 /*
5222  * Allocate and build a list of all syncqs in a stream and the syncq(s)
5223  * associated with the "q" parameter. The resulting list is sorted in a
5224  * canonical order and is free of duplicates.
5225  * Assumes the passed queue is a _RD(q).
5226  */
5227 static sqlist_t *
sqlist_build(queue_t * q,struct stdata * stp,boolean_t do_twist)5228 sqlist_build(queue_t *q, struct stdata *stp, boolean_t do_twist)
5229 {
5230 	sqlist_t *sqlist = sqlist_alloc(stp, KM_SLEEP);
5231 
5232 	/*
5233 	 * start with the current queue/qpair
5234 	 */
5235 	ASSERT(q->q_flag & QREADR);
5236 
5237 	sqlist_insert(sqlist, q->q_syncq);
5238 	sqlist_insert(sqlist, _WR(q)->q_syncq);
5239 
5240 	sqlist_insertall(sqlist, stp->sd_wrq);
5241 	if (do_twist)
5242 		sqlist_insertall(sqlist, stp->sd_mate->sd_wrq);
5243 
5244 	return (sqlist);
5245 }
5246 
5247 static sqlist_t *
sqlist_alloc(struct stdata * stp,int kmflag)5248 sqlist_alloc(struct stdata *stp, int kmflag)
5249 {
5250 	size_t sqlist_size;
5251 	sqlist_t *sqlist;
5252 
5253 	/*
5254 	 * Allocate 2 syncql_t's for each pushed module. Note that
5255 	 * the sqlist_t structure already has 4 syncql_t's built in:
5256 	 * 2 for the stream head, and 2 for the driver/other stream head.
5257 	 */
5258 	sqlist_size = 2 * sizeof (syncql_t) * stp->sd_pushcnt +
5259 	    sizeof (sqlist_t);
5260 	if (STRMATED(stp))
5261 		sqlist_size += 2 * sizeof (syncql_t) * stp->sd_mate->sd_pushcnt;
5262 	sqlist = kmem_alloc(sqlist_size, kmflag);
5263 
5264 	sqlist->sqlist_head = NULL;
5265 	sqlist->sqlist_size = sqlist_size;
5266 	sqlist->sqlist_index = 0;
5267 
5268 	return (sqlist);
5269 }
5270 
5271 /*
5272  * Free the list created by sqlist_alloc()
5273  */
5274 static void
sqlist_free(sqlist_t * sqlist)5275 sqlist_free(sqlist_t *sqlist)
5276 {
5277 	kmem_free(sqlist, sqlist->sqlist_size);
5278 }
5279 
5280 /*
5281  * Prevent any new entries into any syncq in this stream.
5282  * Used by freezestr.
5283  */
5284 void
strblock(queue_t * q)5285 strblock(queue_t *q)
5286 {
5287 	struct stdata	*stp;
5288 	syncql_t	*sql;
5289 	sqlist_t	*sqlist;
5290 
5291 	q = _RD(q);
5292 
5293 	stp = STREAM(q);
5294 	ASSERT(stp != NULL);
5295 
5296 	/*
5297 	 * Get a sorted list with all the duplicates removed containing
5298 	 * all the syncqs referenced by this stream.
5299 	 */
5300 	sqlist = sqlist_build(q, stp, B_FALSE);
5301 	for (sql = sqlist->sqlist_head; sql != NULL; sql = sql->sql_next)
5302 		blocksq(sql->sql_sq, SQ_FROZEN, -1);
5303 	sqlist_free(sqlist);
5304 }
5305 
5306 /*
5307  * Release the block on new entries into this stream
5308  */
5309 void
strunblock(queue_t * q)5310 strunblock(queue_t *q)
5311 {
5312 	struct stdata	*stp;
5313 	syncql_t	*sql;
5314 	sqlist_t	*sqlist;
5315 	int		drain_needed;
5316 
5317 	q = _RD(q);
5318 
5319 	/*
5320 	 * Get a sorted list with all the duplicates removed containing
5321 	 * all the syncqs referenced by this stream.
5322 	 * Have to drop the SQ_FROZEN flag on all the syncqs before
5323 	 * starting to drain them; otherwise the draining might
5324 	 * cause a freezestr in some module on the stream (which
5325 	 * would deadlock).
5326 	 */
5327 	stp = STREAM(q);
5328 	ASSERT(stp != NULL);
5329 	sqlist = sqlist_build(q, stp, B_FALSE);
5330 	drain_needed = 0;
5331 	for (sql = sqlist->sqlist_head; sql != NULL; sql = sql->sql_next)
5332 		drain_needed += dropsq(sql->sql_sq, SQ_FROZEN);
5333 	if (drain_needed) {
5334 		for (sql = sqlist->sqlist_head; sql != NULL;
5335 		    sql = sql->sql_next)
5336 			emptysq(sql->sql_sq);
5337 	}
5338 	sqlist_free(sqlist);
5339 }
5340 
5341 #ifdef DEBUG
5342 static int
qprocsareon(queue_t * rq)5343 qprocsareon(queue_t *rq)
5344 {
5345 	if (rq->q_next == NULL)
5346 		return (0);
5347 	return (_WR(rq->q_next)->q_next == _WR(rq));
5348 }
5349 
5350 int
qclaimed(queue_t * q)5351 qclaimed(queue_t *q)
5352 {
5353 	uint_t count;
5354 
5355 	count = q->q_syncq->sq_count;
5356 	SUM_SQ_PUTCOUNTS(q->q_syncq, count);
5357 	return (count != 0);
5358 }
5359 
5360 /*
5361  * Check if anyone has frozen this stream with freezestr
5362  */
5363 int
frozenstr(queue_t * q)5364 frozenstr(queue_t *q)
5365 {
5366 	return ((q->q_syncq->sq_flags & SQ_FROZEN) != 0);
5367 }
5368 #endif /* DEBUG */
5369 
5370 /*
5371  * Enter a queue.
5372  * Obsoleted interface. Should not be used.
5373  */
5374 void
enterq(queue_t * q)5375 enterq(queue_t *q)
5376 {
5377 	entersq(q->q_syncq, SQ_CALLBACK);
5378 }
5379 
5380 void
leaveq(queue_t * q)5381 leaveq(queue_t *q)
5382 {
5383 	leavesq(q->q_syncq, SQ_CALLBACK);
5384 }
5385 
5386 /*
5387  * Enter a perimeter. c_inner and c_outer specifies which concurrency bits
5388  * to check.
5389  * Wait if SQ_QUEUED is set to preserve ordering between messages and qwriter
5390  * calls and the running of open, close and service procedures.
5391  *
5392  * If c_inner bit is set no need to grab sq_putlocks since we don't care
5393  * if other threads have entered or are entering put entry point.
5394  *
5395  * If c_inner bit is set it might have been possible to use
5396  * sq_putlocks/sq_putcounts instead of SQLOCK/sq_count (e.g. to optimize
5397  * open/close path for IP) but since the count may need to be decremented in
5398  * qwait() we wouldn't know which counter to decrement. Currently counter is
5399  * selected by current cpu_seqid and current CPU can change at any moment. XXX
5400  * in the future we might use curthread id bits to select the counter and this
5401  * would stay constant across routine calls.
5402  */
5403 void
entersq(syncq_t * sq,int entrypoint)5404 entersq(syncq_t *sq, int entrypoint)
5405 {
5406 	uint16_t	count = 0;
5407 	uint16_t	flags;
5408 	uint16_t	waitflags = SQ_STAYAWAY | SQ_EVENTS | SQ_EXCL;
5409 	uint16_t	type;
5410 	uint_t		c_inner = entrypoint & SQ_CI;
5411 	uint_t		c_outer = entrypoint & SQ_CO;
5412 
5413 	/*
5414 	 * Increment ref count to keep closes out of this queue.
5415 	 */
5416 	ASSERT(sq);
5417 	ASSERT(c_inner && c_outer);
5418 	mutex_enter(SQLOCK(sq));
5419 	flags = sq->sq_flags;
5420 	type = sq->sq_type;
5421 	if (!(type & c_inner)) {
5422 		/* Make sure all putcounts now use slowlock. */
5423 		count = sq->sq_count;
5424 		SQ_PUTLOCKS_ENTER(sq);
5425 		SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
5426 		SUM_SQ_PUTCOUNTS(sq, count);
5427 		sq->sq_needexcl++;
5428 		ASSERT(sq->sq_needexcl != 0);	/* wraparound */
5429 		waitflags |= SQ_MESSAGES;
5430 	}
5431 	/*
5432 	 * Wait until we can enter the inner perimeter.
5433 	 * If we want exclusive access we wait until sq_count is 0.
5434 	 * We have to do this before entering the outer perimeter in order
5435 	 * to preserve put/close message ordering.
5436 	 */
5437 	while ((flags & waitflags) || (!(type & c_inner) && count != 0)) {
5438 		sq->sq_flags = flags | SQ_WANTWAKEUP;
5439 		if (!(type & c_inner)) {
5440 			SQ_PUTLOCKS_EXIT(sq);
5441 		}
5442 		cv_wait(&sq->sq_wait, SQLOCK(sq));
5443 		if (!(type & c_inner)) {
5444 			count = sq->sq_count;
5445 			SQ_PUTLOCKS_ENTER(sq);
5446 			SUM_SQ_PUTCOUNTS(sq, count);
5447 		}
5448 		flags = sq->sq_flags;
5449 	}
5450 
5451 	if (!(type & c_inner)) {
5452 		ASSERT(sq->sq_needexcl > 0);
5453 		sq->sq_needexcl--;
5454 		if (sq->sq_needexcl == 0) {
5455 			SQ_PUTCOUNT_SETFAST_LOCKED(sq);
5456 		}
5457 	}
5458 
5459 	/* Check if we need to enter the outer perimeter */
5460 	if (!(type & c_outer)) {
5461 		/*
5462 		 * We have to enter the outer perimeter exclusively before
5463 		 * we can increment sq_count to avoid deadlock. This implies
5464 		 * that we have to re-check sq_flags and sq_count.
5465 		 *
5466 		 * is it possible to have c_inner set when c_outer is not set?
5467 		 */
5468 		if (!(type & c_inner)) {
5469 			SQ_PUTLOCKS_EXIT(sq);
5470 		}
5471 		mutex_exit(SQLOCK(sq));
5472 		outer_enter(sq->sq_outer, SQ_GOAWAY);
5473 		mutex_enter(SQLOCK(sq));
5474 		flags = sq->sq_flags;
5475 		/*
5476 		 * there should be no need to recheck sq_putcounts
5477 		 * because outer_enter() has already waited for them to clear
5478 		 * after setting SQ_WRITER.
5479 		 */
5480 		count = sq->sq_count;
5481 #ifdef DEBUG
5482 		/*
5483 		 * SUMCHECK_SQ_PUTCOUNTS should return the sum instead
5484 		 * of doing an ASSERT internally. Others should do
5485 		 * something like
5486 		 *	 ASSERT(SUMCHECK_SQ_PUTCOUNTS(sq) == 0);
5487 		 * without the need to #ifdef DEBUG it.
5488 		 */
5489 		SUMCHECK_SQ_PUTCOUNTS(sq, 0);
5490 #endif
5491 		while ((flags & (SQ_EXCL|SQ_BLOCKED|SQ_FROZEN)) ||
5492 		    (!(type & c_inner) && count != 0)) {
5493 			sq->sq_flags = flags | SQ_WANTWAKEUP;
5494 			cv_wait(&sq->sq_wait, SQLOCK(sq));
5495 			count = sq->sq_count;
5496 			flags = sq->sq_flags;
5497 		}
5498 	}
5499 
5500 	sq->sq_count++;
5501 	ASSERT(sq->sq_count != 0);	/* Wraparound */
5502 	if (!(type & c_inner)) {
5503 		/* Exclusive entry */
5504 		ASSERT(sq->sq_count == 1);
5505 		sq->sq_flags |= SQ_EXCL;
5506 		if (type & c_outer) {
5507 			SQ_PUTLOCKS_EXIT(sq);
5508 		}
5509 	}
5510 	mutex_exit(SQLOCK(sq));
5511 }
5512 
5513 /*
5514  * Leave a syncq. Announce to framework that closes may proceed.
5515  * c_inner and c_outer specify which concurrency bits to check.
5516  *
5517  * Must never be called from driver or module put entry point.
5518  *
5519  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5520  * sq_putlocks are used.
5521  */
5522 void
leavesq(syncq_t * sq,int entrypoint)5523 leavesq(syncq_t *sq, int entrypoint)
5524 {
5525 	uint16_t	flags;
5526 	uint16_t	type;
5527 	uint_t		c_outer = entrypoint & SQ_CO;
5528 #ifdef DEBUG
5529 	uint_t		c_inner = entrypoint & SQ_CI;
5530 #endif
5531 
5532 	/*
5533 	 * Decrement ref count, drain the syncq if possible, and wake up
5534 	 * any waiting close.
5535 	 */
5536 	ASSERT(sq);
5537 	ASSERT(c_inner && c_outer);
5538 	mutex_enter(SQLOCK(sq));
5539 	flags = sq->sq_flags;
5540 	type = sq->sq_type;
5541 	if (flags & (SQ_QUEUED|SQ_WANTWAKEUP|SQ_WANTEXWAKEUP)) {
5542 
5543 		if (flags & SQ_WANTWAKEUP) {
5544 			flags &= ~SQ_WANTWAKEUP;
5545 			cv_broadcast(&sq->sq_wait);
5546 		}
5547 		if (flags & SQ_WANTEXWAKEUP) {
5548 			flags &= ~SQ_WANTEXWAKEUP;
5549 			cv_broadcast(&sq->sq_exitwait);
5550 		}
5551 
5552 		if ((flags & SQ_QUEUED) && !(flags & SQ_STAYAWAY)) {
5553 			/*
5554 			 * The syncq needs to be drained. "Exit" the syncq
5555 			 * before calling drain_syncq.
5556 			 */
5557 			ASSERT(sq->sq_count != 0);
5558 			sq->sq_count--;
5559 			ASSERT((flags & SQ_EXCL) || (type & c_inner));
5560 			sq->sq_flags = flags & ~SQ_EXCL;
5561 			drain_syncq(sq);
5562 			ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
5563 			/* Check if we need to exit the outer perimeter */
5564 			/* XXX will this ever be true? */
5565 			if (!(type & c_outer))
5566 				outer_exit(sq->sq_outer);
5567 			return;
5568 		}
5569 	}
5570 	ASSERT(sq->sq_count != 0);
5571 	sq->sq_count--;
5572 	ASSERT((flags & SQ_EXCL) || (type & c_inner));
5573 	sq->sq_flags = flags & ~SQ_EXCL;
5574 	mutex_exit(SQLOCK(sq));
5575 
5576 	/* Check if we need to exit the outer perimeter */
5577 	if (!(sq->sq_type & c_outer))
5578 		outer_exit(sq->sq_outer);
5579 }
5580 
5581 /*
5582  * Prevent q_next from changing in this stream by incrementing sq_count.
5583  *
5584  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5585  * sq_putlocks are used.
5586  */
5587 void
claimq(queue_t * qp)5588 claimq(queue_t *qp)
5589 {
5590 	syncq_t	*sq = qp->q_syncq;
5591 
5592 	mutex_enter(SQLOCK(sq));
5593 	sq->sq_count++;
5594 	ASSERT(sq->sq_count != 0);	/* Wraparound */
5595 	mutex_exit(SQLOCK(sq));
5596 }
5597 
5598 /*
5599  * Undo claimq.
5600  *
5601  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5602  * sq_putlocks are used.
5603  */
5604 void
releaseq(queue_t * qp)5605 releaseq(queue_t *qp)
5606 {
5607 	syncq_t	*sq = qp->q_syncq;
5608 	uint16_t flags;
5609 
5610 	mutex_enter(SQLOCK(sq));
5611 	ASSERT(sq->sq_count > 0);
5612 	sq->sq_count--;
5613 
5614 	flags = sq->sq_flags;
5615 	if (flags & (SQ_WANTWAKEUP|SQ_QUEUED)) {
5616 		if (flags & SQ_WANTWAKEUP) {
5617 			flags &= ~SQ_WANTWAKEUP;
5618 			cv_broadcast(&sq->sq_wait);
5619 		}
5620 		sq->sq_flags = flags;
5621 		if ((flags & SQ_QUEUED) && !(flags & (SQ_STAYAWAY|SQ_EXCL))) {
5622 			/*
5623 			 * To prevent potential recursive invocation of
5624 			 * drain_syncq we do not call drain_syncq if count is
5625 			 * non-zero.
5626 			 */
5627 			if (sq->sq_count == 0) {
5628 				drain_syncq(sq);
5629 				return;
5630 			} else
5631 				sqenable(sq);
5632 		}
5633 	}
5634 	mutex_exit(SQLOCK(sq));
5635 }
5636 
5637 /*
5638  * Prevent q_next from changing in this stream by incrementing sd_refcnt.
5639  */
5640 void
claimstr(queue_t * qp)5641 claimstr(queue_t *qp)
5642 {
5643 	struct stdata *stp = STREAM(qp);
5644 
5645 	mutex_enter(&stp->sd_reflock);
5646 	stp->sd_refcnt++;
5647 	ASSERT(stp->sd_refcnt != 0);	/* Wraparound */
5648 	mutex_exit(&stp->sd_reflock);
5649 }
5650 
5651 /*
5652  * Undo claimstr.
5653  */
5654 void
releasestr(queue_t * qp)5655 releasestr(queue_t *qp)
5656 {
5657 	struct stdata *stp = STREAM(qp);
5658 
5659 	mutex_enter(&stp->sd_reflock);
5660 	ASSERT(stp->sd_refcnt != 0);
5661 	if (--stp->sd_refcnt == 0)
5662 		cv_broadcast(&stp->sd_refmonitor);
5663 	mutex_exit(&stp->sd_reflock);
5664 }
5665 
5666 static syncq_t *
new_syncq(void)5667 new_syncq(void)
5668 {
5669 	return (kmem_cache_alloc(syncq_cache, KM_SLEEP));
5670 }
5671 
5672 static void
free_syncq(syncq_t * sq)5673 free_syncq(syncq_t *sq)
5674 {
5675 	ASSERT(sq->sq_head == NULL);
5676 	ASSERT(sq->sq_outer == NULL);
5677 	ASSERT(sq->sq_callbpend == NULL);
5678 	ASSERT((sq->sq_onext == NULL && sq->sq_oprev == NULL) ||
5679 	    (sq->sq_onext == sq && sq->sq_oprev == sq));
5680 
5681 	if (sq->sq_ciputctrl != NULL) {
5682 		ASSERT(sq->sq_nciputctrl == n_ciputctrl - 1);
5683 		SUMCHECK_CIPUTCTRL_COUNTS(sq->sq_ciputctrl,
5684 		    sq->sq_nciputctrl, 0);
5685 		ASSERT(ciputctrl_cache != NULL);
5686 		kmem_cache_free(ciputctrl_cache, sq->sq_ciputctrl);
5687 	}
5688 
5689 	sq->sq_tail = NULL;
5690 	sq->sq_evhead = NULL;
5691 	sq->sq_evtail = NULL;
5692 	sq->sq_ciputctrl = NULL;
5693 	sq->sq_nciputctrl = 0;
5694 	sq->sq_count = 0;
5695 	sq->sq_rmqcount = 0;
5696 	sq->sq_callbflags = 0;
5697 	sq->sq_cancelid = 0;
5698 	sq->sq_next = NULL;
5699 	sq->sq_needexcl = 0;
5700 	sq->sq_svcflags = 0;
5701 	sq->sq_nqueues = 0;
5702 	sq->sq_pri = 0;
5703 	sq->sq_onext = NULL;
5704 	sq->sq_oprev = NULL;
5705 	sq->sq_flags = 0;
5706 	sq->sq_type = 0;
5707 	sq->sq_servcount = 0;
5708 
5709 	kmem_cache_free(syncq_cache, sq);
5710 }
5711 
5712 /* Outer perimeter code */
5713 
5714 /*
5715  * The outer syncq uses the fields and flags in the syncq slightly
5716  * differently from the inner syncqs.
5717  *	sq_count	Incremented when there are pending or running
5718  *			writers at the outer perimeter to prevent the set of
5719  *			inner syncqs that belong to the outer perimeter from
5720  *			changing.
5721  *	sq_head/tail	List of deferred qwriter(OUTER) operations.
5722  *
5723  *	SQ_BLOCKED	Set to prevent traversing of sq_next,sq_prev while
5724  *			inner syncqs are added to or removed from the
5725  *			outer perimeter.
5726  *	SQ_QUEUED	sq_head/tail has messages or events queued.
5727  *
5728  *	SQ_WRITER	A thread is currently traversing all the inner syncqs
5729  *			setting the SQ_WRITER flag.
5730  */
5731 
5732 /*
5733  * Get write access at the outer perimeter.
5734  * Note that read access is done by entersq, putnext, and put by simply
5735  * incrementing sq_count in the inner syncq.
5736  *
5737  * Waits until "flags" is no longer set in the outer to prevent multiple
5738  * threads from having write access at the same time. SQ_WRITER has to be part
5739  * of "flags".
5740  *
5741  * Increases sq_count on the outer syncq to keep away outer_insert/remove
5742  * until the outer_exit is finished.
5743  *
5744  * outer_enter is vulnerable to starvation since it does not prevent new
5745  * threads from entering the inner syncqs while it is waiting for sq_count to
5746  * go to zero.
5747  */
5748 void
outer_enter(syncq_t * outer,uint16_t flags)5749 outer_enter(syncq_t *outer, uint16_t flags)
5750 {
5751 	syncq_t	*sq;
5752 	int	wait_needed;
5753 	uint16_t	count;
5754 
5755 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5756 	    outer->sq_oprev != NULL);
5757 	ASSERT(flags & SQ_WRITER);
5758 
5759 retry:
5760 	mutex_enter(SQLOCK(outer));
5761 	while (outer->sq_flags & flags) {
5762 		outer->sq_flags |= SQ_WANTWAKEUP;
5763 		cv_wait(&outer->sq_wait, SQLOCK(outer));
5764 	}
5765 
5766 	ASSERT(!(outer->sq_flags & SQ_WRITER));
5767 	outer->sq_flags |= SQ_WRITER;
5768 	outer->sq_count++;
5769 	ASSERT(outer->sq_count != 0);	/* wraparound */
5770 	wait_needed = 0;
5771 	/*
5772 	 * Set SQ_WRITER on all the inner syncqs while holding
5773 	 * the SQLOCK on the outer syncq. This ensures that the changing
5774 	 * of SQ_WRITER is atomic under the outer SQLOCK.
5775 	 */
5776 	for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext) {
5777 		mutex_enter(SQLOCK(sq));
5778 		count = sq->sq_count;
5779 		SQ_PUTLOCKS_ENTER(sq);
5780 		sq->sq_flags |= SQ_WRITER;
5781 		SUM_SQ_PUTCOUNTS(sq, count);
5782 		if (count != 0)
5783 			wait_needed = 1;
5784 		SQ_PUTLOCKS_EXIT(sq);
5785 		mutex_exit(SQLOCK(sq));
5786 	}
5787 	mutex_exit(SQLOCK(outer));
5788 
5789 	/*
5790 	 * Get everybody out of the syncqs sequentially.
5791 	 * Note that we don't actually need to acquire the PUTLOCKS, since
5792 	 * we have already cleared the fastbit, and set QWRITER.  By
5793 	 * definition, the count can not increase since putnext will
5794 	 * take the slowlock path (and the purpose of acquiring the
5795 	 * putlocks was to make sure it didn't increase while we were
5796 	 * waiting).
5797 	 *
5798 	 * Note that we still acquire the PUTLOCKS to be safe.
5799 	 */
5800 	if (wait_needed) {
5801 		for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext) {
5802 			mutex_enter(SQLOCK(sq));
5803 			count = sq->sq_count;
5804 			SQ_PUTLOCKS_ENTER(sq);
5805 			SUM_SQ_PUTCOUNTS(sq, count);
5806 			while (count != 0) {
5807 				sq->sq_flags |= SQ_WANTWAKEUP;
5808 				SQ_PUTLOCKS_EXIT(sq);
5809 				cv_wait(&sq->sq_wait, SQLOCK(sq));
5810 				count = sq->sq_count;
5811 				SQ_PUTLOCKS_ENTER(sq);
5812 				SUM_SQ_PUTCOUNTS(sq, count);
5813 			}
5814 			SQ_PUTLOCKS_EXIT(sq);
5815 			mutex_exit(SQLOCK(sq));
5816 		}
5817 		/*
5818 		 * Verify that none of the flags got set while we
5819 		 * were waiting for the sq_counts to drop.
5820 		 * If this happens we exit and retry entering the
5821 		 * outer perimeter.
5822 		 */
5823 		mutex_enter(SQLOCK(outer));
5824 		if (outer->sq_flags & (flags & ~SQ_WRITER)) {
5825 			mutex_exit(SQLOCK(outer));
5826 			outer_exit(outer);
5827 			goto retry;
5828 		}
5829 		mutex_exit(SQLOCK(outer));
5830 	}
5831 }
5832 
5833 /*
5834  * Drop the write access at the outer perimeter.
5835  * Read access is dropped implicitly (by putnext, put, and leavesq) by
5836  * decrementing sq_count.
5837  */
5838 void
outer_exit(syncq_t * outer)5839 outer_exit(syncq_t *outer)
5840 {
5841 	syncq_t	*sq;
5842 	int	 drain_needed;
5843 	uint16_t flags;
5844 
5845 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5846 	    outer->sq_oprev != NULL);
5847 	ASSERT(MUTEX_NOT_HELD(SQLOCK(outer)));
5848 
5849 	/*
5850 	 * Atomically (from the perspective of threads calling become_writer)
5851 	 * drop the write access at the outer perimeter by holding
5852 	 * SQLOCK(outer) across all the dropsq calls and the resetting of
5853 	 * SQ_WRITER.
5854 	 * This defines a locking order between the outer perimeter
5855 	 * SQLOCK and the inner perimeter SQLOCKs.
5856 	 */
5857 	mutex_enter(SQLOCK(outer));
5858 	flags = outer->sq_flags;
5859 	ASSERT(outer->sq_flags & SQ_WRITER);
5860 	if (flags & SQ_QUEUED) {
5861 		write_now(outer);
5862 		flags = outer->sq_flags;
5863 	}
5864 
5865 	/*
5866 	 * sq_onext is stable since sq_count has not yet been decreased.
5867 	 * Reset the SQ_WRITER flags in all syncqs.
5868 	 * After dropping SQ_WRITER on the outer syncq we empty all the
5869 	 * inner syncqs.
5870 	 */
5871 	drain_needed = 0;
5872 	for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext)
5873 		drain_needed += dropsq(sq, SQ_WRITER);
5874 	ASSERT(!(outer->sq_flags & SQ_QUEUED));
5875 	flags &= ~SQ_WRITER;
5876 	if (drain_needed) {
5877 		outer->sq_flags = flags;
5878 		mutex_exit(SQLOCK(outer));
5879 		for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext)
5880 			emptysq(sq);
5881 		mutex_enter(SQLOCK(outer));
5882 		flags = outer->sq_flags;
5883 	}
5884 	if (flags & SQ_WANTWAKEUP) {
5885 		flags &= ~SQ_WANTWAKEUP;
5886 		cv_broadcast(&outer->sq_wait);
5887 	}
5888 	outer->sq_flags = flags;
5889 	ASSERT(outer->sq_count > 0);
5890 	outer->sq_count--;
5891 	mutex_exit(SQLOCK(outer));
5892 }
5893 
5894 /*
5895  * Add another syncq to an outer perimeter.
5896  * Block out all other access to the outer perimeter while it is being
5897  * changed using blocksq.
5898  * Assumes that the caller has *not* done an outer_enter.
5899  *
5900  * Vulnerable to starvation in blocksq.
5901  */
5902 static void
outer_insert(syncq_t * outer,syncq_t * sq)5903 outer_insert(syncq_t *outer, syncq_t *sq)
5904 {
5905 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5906 	    outer->sq_oprev != NULL);
5907 	ASSERT(sq->sq_outer == NULL && sq->sq_onext == NULL &&
5908 	    sq->sq_oprev == NULL);	/* Can't be in an outer perimeter */
5909 
5910 	/* Get exclusive access to the outer perimeter list */
5911 	blocksq(outer, SQ_BLOCKED, 0);
5912 	ASSERT(outer->sq_flags & SQ_BLOCKED);
5913 	ASSERT(!(outer->sq_flags & SQ_WRITER));
5914 
5915 	mutex_enter(SQLOCK(sq));
5916 	sq->sq_outer = outer;
5917 	outer->sq_onext->sq_oprev = sq;
5918 	sq->sq_onext = outer->sq_onext;
5919 	outer->sq_onext = sq;
5920 	sq->sq_oprev = outer;
5921 	mutex_exit(SQLOCK(sq));
5922 	unblocksq(outer, SQ_BLOCKED, 1);
5923 }
5924 
5925 /*
5926  * Remove a syncq from an outer perimeter.
5927  * Block out all other access to the outer perimeter while it is being
5928  * changed using blocksq.
5929  * Assumes that the caller has *not* done an outer_enter.
5930  *
5931  * Vulnerable to starvation in blocksq.
5932  */
5933 static void
outer_remove(syncq_t * outer,syncq_t * sq)5934 outer_remove(syncq_t *outer, syncq_t *sq)
5935 {
5936 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5937 	    outer->sq_oprev != NULL);
5938 	ASSERT(sq->sq_outer == outer);
5939 
5940 	/* Get exclusive access to the outer perimeter list */
5941 	blocksq(outer, SQ_BLOCKED, 0);
5942 	ASSERT(outer->sq_flags & SQ_BLOCKED);
5943 	ASSERT(!(outer->sq_flags & SQ_WRITER));
5944 
5945 	mutex_enter(SQLOCK(sq));
5946 	sq->sq_outer = NULL;
5947 	sq->sq_onext->sq_oprev = sq->sq_oprev;
5948 	sq->sq_oprev->sq_onext = sq->sq_onext;
5949 	sq->sq_oprev = sq->sq_onext = NULL;
5950 	mutex_exit(SQLOCK(sq));
5951 	unblocksq(outer, SQ_BLOCKED, 1);
5952 }
5953 
5954 /*
5955  * Queue a deferred qwriter(OUTER) callback for this outer perimeter.
5956  * If this is the first callback for this outer perimeter then add
5957  * this outer perimeter to the list of outer perimeters that
5958  * the qwriter_outer_thread will process.
5959  *
5960  * Increments sq_count in the outer syncq to prevent the membership
5961  * of the outer perimeter (in terms of inner syncqs) to change while
5962  * the callback is pending.
5963  */
5964 static void
queue_writer(syncq_t * outer,void (* func)(),queue_t * q,mblk_t * mp)5965 queue_writer(syncq_t *outer, void (*func)(), queue_t *q, mblk_t *mp)
5966 {
5967 	ASSERT(MUTEX_HELD(SQLOCK(outer)));
5968 
5969 	mp->b_prev = (mblk_t *)func;
5970 	mp->b_queue = q;
5971 	mp->b_next = NULL;
5972 	outer->sq_count++;	/* Decremented when dequeued */
5973 	ASSERT(outer->sq_count != 0);	/* Wraparound */
5974 	if (outer->sq_evhead == NULL) {
5975 		/* First message. */
5976 		outer->sq_evhead = outer->sq_evtail = mp;
5977 		outer->sq_flags |= SQ_EVENTS;
5978 		mutex_exit(SQLOCK(outer));
5979 		STRSTAT(qwr_outer);
5980 		(void) taskq_dispatch(streams_taskq,
5981 		    (task_func_t *)qwriter_outer_service, outer, TQ_SLEEP);
5982 	} else {
5983 		ASSERT(outer->sq_flags & SQ_EVENTS);
5984 		outer->sq_evtail->b_next = mp;
5985 		outer->sq_evtail = mp;
5986 		mutex_exit(SQLOCK(outer));
5987 	}
5988 }
5989 
5990 /*
5991  * Try and upgrade to write access at the outer perimeter. If this can
5992  * not be done without blocking then queue the callback to be done
5993  * by the qwriter_outer_thread.
5994  *
5995  * This routine can only be called from put or service procedures plus
5996  * asynchronous callback routines that have properly entered the queue (with
5997  * entersq). Thus qwriter(OUTER) assumes the caller has one claim on the syncq
5998  * associated with q.
5999  */
6000 void
qwriter_outer(queue_t * q,mblk_t * mp,void (* func)())6001 qwriter_outer(queue_t *q, mblk_t *mp, void (*func)())
6002 {
6003 	syncq_t	*osq, *sq, *outer;
6004 	int	failed;
6005 	uint16_t flags;
6006 
6007 	osq = q->q_syncq;
6008 	outer = osq->sq_outer;
6009 	if (outer == NULL)
6010 		panic("qwriter(PERIM_OUTER): no outer perimeter");
6011 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
6012 	    outer->sq_oprev != NULL);
6013 
6014 	mutex_enter(SQLOCK(outer));
6015 	flags = outer->sq_flags;
6016 	/*
6017 	 * If some thread is traversing sq_next, or if we are blocked by
6018 	 * outer_insert or outer_remove, or if the we already have queued
6019 	 * callbacks, then queue this callback for later processing.
6020 	 *
6021 	 * Also queue the qwriter for an interrupt thread in order
6022 	 * to reduce the time spent running at high IPL.
6023 	 * to identify there are events.
6024 	 */
6025 	if ((flags & SQ_GOAWAY) || (curthread->t_pri >= kpreemptpri)) {
6026 		/*
6027 		 * Queue the become_writer request.
6028 		 * The queueing is atomic under SQLOCK(outer) in order
6029 		 * to synchronize with outer_exit.
6030 		 * queue_writer will drop the outer SQLOCK
6031 		 */
6032 		if (flags & SQ_BLOCKED) {
6033 			/* Must set SQ_WRITER on inner perimeter */
6034 			mutex_enter(SQLOCK(osq));
6035 			osq->sq_flags |= SQ_WRITER;
6036 			mutex_exit(SQLOCK(osq));
6037 		} else {
6038 			if (!(flags & SQ_WRITER)) {
6039 				/*
6040 				 * The outer could have been SQ_BLOCKED thus
6041 				 * SQ_WRITER might not be set on the inner.
6042 				 */
6043 				mutex_enter(SQLOCK(osq));
6044 				osq->sq_flags |= SQ_WRITER;
6045 				mutex_exit(SQLOCK(osq));
6046 			}
6047 			ASSERT(osq->sq_flags & SQ_WRITER);
6048 		}
6049 		queue_writer(outer, func, q, mp);
6050 		return;
6051 	}
6052 	/*
6053 	 * We are half-way to exclusive access to the outer perimeter.
6054 	 * Prevent any outer_enter, qwriter(OUTER), or outer_insert/remove
6055 	 * while the inner syncqs are traversed.
6056 	 */
6057 	outer->sq_count++;
6058 	ASSERT(outer->sq_count != 0);	/* wraparound */
6059 	flags |= SQ_WRITER;
6060 	/*
6061 	 * Check if we can run the function immediately. Mark all
6062 	 * syncqs with the writer flag to prevent new entries into
6063 	 * put and service procedures.
6064 	 *
6065 	 * Set SQ_WRITER on all the inner syncqs while holding
6066 	 * the SQLOCK on the outer syncq. This ensures that the changing
6067 	 * of SQ_WRITER is atomic under the outer SQLOCK.
6068 	 */
6069 	failed = 0;
6070 	for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext) {
6071 		uint16_t count;
6072 		uint_t	maxcnt = (sq == osq) ? 1 : 0;
6073 
6074 		mutex_enter(SQLOCK(sq));
6075 		count = sq->sq_count;
6076 		SQ_PUTLOCKS_ENTER(sq);
6077 		SUM_SQ_PUTCOUNTS(sq, count);
6078 		if (sq->sq_count > maxcnt)
6079 			failed = 1;
6080 		sq->sq_flags |= SQ_WRITER;
6081 		SQ_PUTLOCKS_EXIT(sq);
6082 		mutex_exit(SQLOCK(sq));
6083 	}
6084 	if (failed) {
6085 		/*
6086 		 * Some other thread has a read claim on the outer perimeter.
6087 		 * Queue the callback for deferred processing.
6088 		 *
6089 		 * queue_writer will set SQ_QUEUED before we drop SQ_WRITER
6090 		 * so that other qwriter(OUTER) calls will queue their
6091 		 * callbacks as well. queue_writer increments sq_count so we
6092 		 * decrement to compensate for the our increment.
6093 		 *
6094 		 * Dropping SQ_WRITER enables the writer thread to work
6095 		 * on this outer perimeter.
6096 		 */
6097 		outer->sq_flags = flags;
6098 		queue_writer(outer, func, q, mp);
6099 		/* queue_writer dropper the lock */
6100 		mutex_enter(SQLOCK(outer));
6101 		ASSERT(outer->sq_count > 0);
6102 		outer->sq_count--;
6103 		ASSERT(outer->sq_flags & SQ_WRITER);
6104 		flags = outer->sq_flags;
6105 		flags &= ~SQ_WRITER;
6106 		if (flags & SQ_WANTWAKEUP) {
6107 			flags &= ~SQ_WANTWAKEUP;
6108 			cv_broadcast(&outer->sq_wait);
6109 		}
6110 		outer->sq_flags = flags;
6111 		mutex_exit(SQLOCK(outer));
6112 		return;
6113 	} else {
6114 		outer->sq_flags = flags;
6115 		mutex_exit(SQLOCK(outer));
6116 	}
6117 
6118 	/* Can run it immediately */
6119 	(*func)(q, mp);
6120 
6121 	outer_exit(outer);
6122 }
6123 
6124 /*
6125  * Dequeue all writer callbacks from the outer perimeter and run them.
6126  */
6127 static void
write_now(syncq_t * outer)6128 write_now(syncq_t *outer)
6129 {
6130 	mblk_t		*mp;
6131 	queue_t		*q;
6132 	void	(*func)();
6133 
6134 	ASSERT(MUTEX_HELD(SQLOCK(outer)));
6135 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
6136 	    outer->sq_oprev != NULL);
6137 	while ((mp = outer->sq_evhead) != NULL) {
6138 		/*
6139 		 * queues cannot be placed on the queuelist on the outer
6140 		 * perimeter.
6141 		 */
6142 		ASSERT(!(outer->sq_flags & SQ_MESSAGES));
6143 		ASSERT((outer->sq_flags & SQ_EVENTS));
6144 
6145 		outer->sq_evhead = mp->b_next;
6146 		if (outer->sq_evhead == NULL) {
6147 			outer->sq_evtail = NULL;
6148 			outer->sq_flags &= ~SQ_EVENTS;
6149 		}
6150 		ASSERT(outer->sq_count != 0);
6151 		outer->sq_count--;	/* Incremented when enqueued. */
6152 		mutex_exit(SQLOCK(outer));
6153 		/*
6154 		 * Drop the message if the queue is closing.
6155 		 * Make sure that the queue is "claimed" when the callback
6156 		 * is run in order to satisfy various ASSERTs.
6157 		 */
6158 		q = mp->b_queue;
6159 		func = (void (*)())mp->b_prev;
6160 		ASSERT(func != NULL);
6161 		mp->b_next = mp->b_prev = NULL;
6162 		if (q->q_flag & QWCLOSE) {
6163 			freemsg(mp);
6164 		} else {
6165 			claimq(q);
6166 			(*func)(q, mp);
6167 			releaseq(q);
6168 		}
6169 		mutex_enter(SQLOCK(outer));
6170 	}
6171 	ASSERT(MUTEX_HELD(SQLOCK(outer)));
6172 }
6173 
6174 /*
6175  * The list of messages on the inner syncq is effectively hashed
6176  * by destination queue.  These destination queues are doubly
6177  * linked lists (hopefully) in priority order.  Messages are then
6178  * put on the queue referenced by the q_sqhead/q_sqtail elements.
6179  * Additional messages are linked together by the b_next/b_prev
6180  * elements in the mblk, with (similar to putq()) the first message
6181  * having a NULL b_prev and the last message having a NULL b_next.
6182  *
6183  * Events, such as qwriter callbacks, are put onto a list in FIFO
6184  * order referenced by sq_evhead, and sq_evtail.  This is a singly
6185  * linked list, and messages here MUST be processed in the order queued.
6186  */
6187 
6188 /*
6189  * Run the events on the syncq event list (sq_evhead).
6190  * Assumes there is only one claim on the syncq, it is
6191  * already exclusive (SQ_EXCL set), and the SQLOCK held.
6192  * Messages here are processed in order, with the SQ_EXCL bit
6193  * held all the way through till the last message is processed.
6194  */
6195 void
sq_run_events(syncq_t * sq)6196 sq_run_events(syncq_t *sq)
6197 {
6198 	mblk_t		*bp;
6199 	queue_t		*qp;
6200 	uint16_t	flags = sq->sq_flags;
6201 	void		(*func)();
6202 
6203 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6204 	ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6205 	    sq->sq_oprev == NULL) ||
6206 	    (sq->sq_outer != NULL && sq->sq_onext != NULL &&
6207 	    sq->sq_oprev != NULL));
6208 
6209 	ASSERT(flags & SQ_EXCL);
6210 	ASSERT(sq->sq_count == 1);
6211 
6212 	/*
6213 	 * We need to process all of the events on this list.  It
6214 	 * is possible that new events will be added while we are
6215 	 * away processing a callback, so on every loop, we start
6216 	 * back at the beginning of the list.
6217 	 */
6218 	/*
6219 	 * We have to reaccess sq_evhead since there is a
6220 	 * possibility of a new entry while we were running
6221 	 * the callback.
6222 	 */
6223 	for (bp = sq->sq_evhead; bp != NULL; bp = sq->sq_evhead) {
6224 		ASSERT(bp->b_queue->q_syncq == sq);
6225 		ASSERT(sq->sq_flags & SQ_EVENTS);
6226 
6227 		qp = bp->b_queue;
6228 		func = (void (*)())bp->b_prev;
6229 		ASSERT(func != NULL);
6230 
6231 		/*
6232 		 * Messages from the event queue must be taken off in
6233 		 * FIFO order.
6234 		 */
6235 		ASSERT(sq->sq_evhead == bp);
6236 		sq->sq_evhead = bp->b_next;
6237 
6238 		if (bp->b_next == NULL) {
6239 			/* Deleting last */
6240 			ASSERT(sq->sq_evtail == bp);
6241 			sq->sq_evtail = NULL;
6242 			sq->sq_flags &= ~SQ_EVENTS;
6243 		}
6244 		bp->b_prev = bp->b_next = NULL;
6245 		ASSERT(bp->b_datap->db_ref != 0);
6246 
6247 		mutex_exit(SQLOCK(sq));
6248 
6249 		(*func)(qp, bp);
6250 
6251 		mutex_enter(SQLOCK(sq));
6252 		/*
6253 		 * re-read the flags, since they could have changed.
6254 		 */
6255 		flags = sq->sq_flags;
6256 		ASSERT(flags & SQ_EXCL);
6257 	}
6258 	ASSERT(sq->sq_evhead == NULL && sq->sq_evtail == NULL);
6259 	ASSERT(!(sq->sq_flags & SQ_EVENTS));
6260 
6261 	if (flags & SQ_WANTWAKEUP) {
6262 		flags &= ~SQ_WANTWAKEUP;
6263 		cv_broadcast(&sq->sq_wait);
6264 	}
6265 	if (flags & SQ_WANTEXWAKEUP) {
6266 		flags &= ~SQ_WANTEXWAKEUP;
6267 		cv_broadcast(&sq->sq_exitwait);
6268 	}
6269 	sq->sq_flags = flags;
6270 }
6271 
6272 /*
6273  * Put messages on the event list.
6274  * If we can go exclusive now, do so and process the event list, otherwise
6275  * let the last claim service this list (or wake the sqthread).
6276  * This procedure assumes SQLOCK is held.  To run the event list, it
6277  * must be called with no claims.
6278  */
6279 static void
sqfill_events(syncq_t * sq,queue_t * q,mblk_t * mp,void (* func)())6280 sqfill_events(syncq_t *sq, queue_t *q, mblk_t *mp, void (*func)())
6281 {
6282 	uint16_t count;
6283 
6284 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6285 	ASSERT(func != NULL);
6286 
6287 	/*
6288 	 * This is a callback.  Add it to the list of callbacks
6289 	 * and see about upgrading.
6290 	 */
6291 	mp->b_prev = (mblk_t *)func;
6292 	mp->b_queue = q;
6293 	mp->b_next = NULL;
6294 	if (sq->sq_evhead == NULL) {
6295 		sq->sq_evhead = sq->sq_evtail = mp;
6296 		sq->sq_flags |= SQ_EVENTS;
6297 	} else {
6298 		ASSERT(sq->sq_evtail != NULL);
6299 		ASSERT(sq->sq_evtail->b_next == NULL);
6300 		ASSERT(sq->sq_flags & SQ_EVENTS);
6301 		sq->sq_evtail->b_next = mp;
6302 		sq->sq_evtail = mp;
6303 	}
6304 	/*
6305 	 * We have set SQ_EVENTS, so threads will have to
6306 	 * unwind out of the perimeter, and new entries will
6307 	 * not grab a putlock.  But we still need to know
6308 	 * how many threads have already made a claim to the
6309 	 * syncq, so grab the putlocks, and sum the counts.
6310 	 * If there are no claims on the syncq, we can upgrade
6311 	 * to exclusive, and run the event list.
6312 	 * NOTE: We hold the SQLOCK, so we can just grab the
6313 	 * putlocks.
6314 	 */
6315 	count = sq->sq_count;
6316 	SQ_PUTLOCKS_ENTER(sq);
6317 	SUM_SQ_PUTCOUNTS(sq, count);
6318 	/*
6319 	 * We have no claim, so we need to check if there
6320 	 * are no others, then we can upgrade.
6321 	 */
6322 	/*
6323 	 * There are currently no claims on
6324 	 * the syncq by this thread (at least on this entry). The thread who has
6325 	 * the claim should drain syncq.
6326 	 */
6327 	if (count > 0) {
6328 		/*
6329 		 * Can't upgrade - other threads inside.
6330 		 */
6331 		SQ_PUTLOCKS_EXIT(sq);
6332 		mutex_exit(SQLOCK(sq));
6333 		return;
6334 	}
6335 	/*
6336 	 * Need to set SQ_EXCL and make a claim on the syncq.
6337 	 */
6338 	ASSERT((sq->sq_flags & SQ_EXCL) == 0);
6339 	sq->sq_flags |= SQ_EXCL;
6340 	ASSERT(sq->sq_count == 0);
6341 	sq->sq_count++;
6342 	SQ_PUTLOCKS_EXIT(sq);
6343 
6344 	/* Process the events list */
6345 	sq_run_events(sq);
6346 
6347 	/*
6348 	 * Release our claim...
6349 	 */
6350 	sq->sq_count--;
6351 
6352 	/*
6353 	 * And release SQ_EXCL.
6354 	 * We don't need to acquire the putlocks to release
6355 	 * SQ_EXCL, since we are exclusive, and hold the SQLOCK.
6356 	 */
6357 	sq->sq_flags &= ~SQ_EXCL;
6358 
6359 	/*
6360 	 * sq_run_events should have released SQ_EXCL
6361 	 */
6362 	ASSERT(!(sq->sq_flags & SQ_EXCL));
6363 
6364 	/*
6365 	 * If anything happened while we were running the
6366 	 * events (or was there before), we need to process
6367 	 * them now.  We shouldn't be exclusive sine we
6368 	 * released the perimeter above (plus, we asserted
6369 	 * for it).
6370 	 */
6371 	if (!(sq->sq_flags & SQ_STAYAWAY) && (sq->sq_flags & SQ_QUEUED))
6372 		drain_syncq(sq);
6373 	else
6374 		mutex_exit(SQLOCK(sq));
6375 }
6376 
6377 /*
6378  * Perform delayed processing. The caller has to make sure that it is safe
6379  * to enter the syncq (e.g. by checking that none of the SQ_STAYAWAY bits are
6380  * set).
6381  *
6382  * Assume that the caller has NO claims on the syncq.  However, a claim
6383  * on the syncq does not indicate that a thread is draining the syncq.
6384  * There may be more claims on the syncq than there are threads draining
6385  * (i.e.  #_threads_draining <= sq_count)
6386  *
6387  * drain_syncq has to terminate when one of the SQ_STAYAWAY bits gets set
6388  * in order to preserve qwriter(OUTER) ordering constraints.
6389  *
6390  * sq_putcount only needs to be checked when dispatching the queued
6391  * writer call for CIPUT sync queue, but this is handled in sq_run_events.
6392  */
6393 void
drain_syncq(syncq_t * sq)6394 drain_syncq(syncq_t *sq)
6395 {
6396 	queue_t		*qp;
6397 	uint16_t	count;
6398 	uint16_t	type = sq->sq_type;
6399 	uint16_t	flags = sq->sq_flags;
6400 	boolean_t	bg_service = sq->sq_svcflags & SQ_SERVICE;
6401 
6402 	TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_START,
6403 	    "drain_syncq start:%p", sq);
6404 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6405 	ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6406 	    sq->sq_oprev == NULL) ||
6407 	    (sq->sq_outer != NULL && sq->sq_onext != NULL &&
6408 	    sq->sq_oprev != NULL));
6409 
6410 	/*
6411 	 * Drop SQ_SERVICE flag.
6412 	 */
6413 	if (bg_service)
6414 		sq->sq_svcflags &= ~SQ_SERVICE;
6415 
6416 	/*
6417 	 * If SQ_EXCL is set, someone else is processing this syncq - let them
6418 	 * finish the job.
6419 	 */
6420 	if (flags & SQ_EXCL) {
6421 		if (bg_service) {
6422 			ASSERT(sq->sq_servcount != 0);
6423 			sq->sq_servcount--;
6424 		}
6425 		mutex_exit(SQLOCK(sq));
6426 		return;
6427 	}
6428 
6429 	/*
6430 	 * This routine can be called by a background thread if
6431 	 * it was scheduled by a hi-priority thread.  SO, if there are
6432 	 * NOT messages queued, return (remember, we have the SQLOCK,
6433 	 * and it cannot change until we release it). Wakeup any waiters also.
6434 	 */
6435 	if (!(flags & SQ_QUEUED)) {
6436 		if (flags & SQ_WANTWAKEUP) {
6437 			flags &= ~SQ_WANTWAKEUP;
6438 			cv_broadcast(&sq->sq_wait);
6439 		}
6440 		if (flags & SQ_WANTEXWAKEUP) {
6441 			flags &= ~SQ_WANTEXWAKEUP;
6442 			cv_broadcast(&sq->sq_exitwait);
6443 		}
6444 		sq->sq_flags = flags;
6445 		if (bg_service) {
6446 			ASSERT(sq->sq_servcount != 0);
6447 			sq->sq_servcount--;
6448 		}
6449 		mutex_exit(SQLOCK(sq));
6450 		return;
6451 	}
6452 
6453 	/*
6454 	 * If this is not a concurrent put perimeter, we need to
6455 	 * become exclusive to drain.  Also, if not CIPUT, we would
6456 	 * not have acquired a putlock, so we don't need to check
6457 	 * the putcounts.  If not entering with a claim, we test
6458 	 * for sq_count == 0.
6459 	 */
6460 	type = sq->sq_type;
6461 	if (!(type & SQ_CIPUT)) {
6462 		if (sq->sq_count > 1) {
6463 			if (bg_service) {
6464 				ASSERT(sq->sq_servcount != 0);
6465 				sq->sq_servcount--;
6466 			}
6467 			mutex_exit(SQLOCK(sq));
6468 			return;
6469 		}
6470 		sq->sq_flags |= SQ_EXCL;
6471 	}
6472 
6473 	/*
6474 	 * This is where we make a claim to the syncq.
6475 	 * This can either be done by incrementing a putlock, or
6476 	 * the sq_count.  But since we already have the SQLOCK
6477 	 * here, we just bump the sq_count.
6478 	 *
6479 	 * Note that after we make a claim, we need to let the code
6480 	 * fall through to the end of this routine to clean itself
6481 	 * up.  A return in the while loop will put the syncq in a
6482 	 * very bad state.
6483 	 */
6484 	sq->sq_count++;
6485 	ASSERT(sq->sq_count != 0);	/* wraparound */
6486 
6487 	while ((flags = sq->sq_flags) & SQ_QUEUED) {
6488 		/*
6489 		 * If we are told to stayaway or went exclusive,
6490 		 * we are done.
6491 		 */
6492 		if (flags & (SQ_STAYAWAY)) {
6493 			break;
6494 		}
6495 
6496 		/*
6497 		 * If there are events to run, do so.
6498 		 * We have one claim to the syncq, so if there are
6499 		 * more than one, other threads are running.
6500 		 */
6501 		if (sq->sq_evhead != NULL) {
6502 			ASSERT(sq->sq_flags & SQ_EVENTS);
6503 
6504 			count = sq->sq_count;
6505 			SQ_PUTLOCKS_ENTER(sq);
6506 			SUM_SQ_PUTCOUNTS(sq, count);
6507 			if (count > 1) {
6508 				SQ_PUTLOCKS_EXIT(sq);
6509 				/* Can't upgrade - other threads inside */
6510 				break;
6511 			}
6512 			ASSERT((flags & SQ_EXCL) == 0);
6513 			sq->sq_flags = flags | SQ_EXCL;
6514 			SQ_PUTLOCKS_EXIT(sq);
6515 			/*
6516 			 * we have the only claim, run the events,
6517 			 * sq_run_events will clear the SQ_EXCL flag.
6518 			 */
6519 			sq_run_events(sq);
6520 
6521 			/*
6522 			 * If this is a CIPUT perimeter, we need
6523 			 * to drop the SQ_EXCL flag so we can properly
6524 			 * continue draining the syncq.
6525 			 */
6526 			if (type & SQ_CIPUT) {
6527 				ASSERT(sq->sq_flags & SQ_EXCL);
6528 				sq->sq_flags &= ~SQ_EXCL;
6529 			}
6530 
6531 			/*
6532 			 * And go back to the beginning just in case
6533 			 * anything changed while we were away.
6534 			 */
6535 			ASSERT((sq->sq_flags & SQ_EXCL) || (type & SQ_CIPUT));
6536 			continue;
6537 		}
6538 
6539 		ASSERT(sq->sq_evhead == NULL);
6540 		ASSERT(!(sq->sq_flags & SQ_EVENTS));
6541 
6542 		/*
6543 		 * Find the queue that is not draining.
6544 		 *
6545 		 * q_draining is protected by QLOCK which we do not hold.
6546 		 * But if it was set, then a thread was draining, and if it gets
6547 		 * cleared, then it was because the thread has successfully
6548 		 * drained the syncq, or a GOAWAY state occurred. For the GOAWAY
6549 		 * state to happen, a thread needs the SQLOCK which we hold, and
6550 		 * if there was such a flag, we would have already seen it.
6551 		 */
6552 
6553 		for (qp = sq->sq_head;
6554 		    qp != NULL && (qp->q_draining ||
6555 		    (qp->q_sqflags & Q_SQDRAINING));
6556 		    qp = qp->q_sqnext)
6557 			;
6558 
6559 		if (qp == NULL)
6560 			break;
6561 
6562 		/*
6563 		 * We have a queue to work on, and we hold the
6564 		 * SQLOCK and one claim, call qdrain_syncq.
6565 		 * This means we need to release the SQLOCK and
6566 		 * acquire the QLOCK (OK since we have a claim).
6567 		 * Note that qdrain_syncq will actually dequeue
6568 		 * this queue from the sq_head list when it is
6569 		 * convinced all the work is done and release
6570 		 * the QLOCK before returning.
6571 		 */
6572 		qp->q_sqflags |= Q_SQDRAINING;
6573 		mutex_exit(SQLOCK(sq));
6574 		mutex_enter(QLOCK(qp));
6575 		qdrain_syncq(sq, qp);
6576 		mutex_enter(SQLOCK(sq));
6577 
6578 		/* The queue is drained */
6579 		ASSERT(qp->q_sqflags & Q_SQDRAINING);
6580 		qp->q_sqflags &= ~Q_SQDRAINING;
6581 		/*
6582 		 * NOTE: After this point qp should not be used since it may be
6583 		 * closed.
6584 		 */
6585 	}
6586 
6587 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6588 	flags = sq->sq_flags;
6589 
6590 	/*
6591 	 * sq->sq_head cannot change because we hold the
6592 	 * sqlock. However, a thread CAN decide that it is no longer
6593 	 * going to drain that queue.  However, this should be due to
6594 	 * a GOAWAY state, and we should see that here.
6595 	 *
6596 	 * This loop is not very efficient. One solution may be adding a second
6597 	 * pointer to the "draining" queue, but it is difficult to do when
6598 	 * queues are inserted in the middle due to priority ordering. Another
6599 	 * possibility is to yank the queue out of the sq list and put it onto
6600 	 * the "draining list" and then put it back if it can't be drained.
6601 	 */
6602 
6603 	ASSERT((sq->sq_head == NULL) || (flags & SQ_GOAWAY) ||
6604 	    (type & SQ_CI) || sq->sq_head->q_draining);
6605 
6606 	/* Drop SQ_EXCL for non-CIPUT perimeters */
6607 	if (!(type & SQ_CIPUT))
6608 		flags &= ~SQ_EXCL;
6609 	ASSERT((flags & SQ_EXCL) == 0);
6610 
6611 	/* Wake up any waiters. */
6612 	if (flags & SQ_WANTWAKEUP) {
6613 		flags &= ~SQ_WANTWAKEUP;
6614 		cv_broadcast(&sq->sq_wait);
6615 	}
6616 	if (flags & SQ_WANTEXWAKEUP) {
6617 		flags &= ~SQ_WANTEXWAKEUP;
6618 		cv_broadcast(&sq->sq_exitwait);
6619 	}
6620 	sq->sq_flags = flags;
6621 
6622 	ASSERT(sq->sq_count != 0);
6623 	/* Release our claim. */
6624 	sq->sq_count--;
6625 
6626 	if (bg_service) {
6627 		ASSERT(sq->sq_servcount != 0);
6628 		sq->sq_servcount--;
6629 	}
6630 
6631 	mutex_exit(SQLOCK(sq));
6632 
6633 	TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_END,
6634 	    "drain_syncq end:%p", sq);
6635 }
6636 
6637 
6638 /*
6639  *
6640  * qdrain_syncq can be called (currently) from only one of two places:
6641  *	drain_syncq
6642  *	putnext  (or some variation of it).
6643  * and eventually
6644  *	qwait(_sig)
6645  *
6646  * If called from drain_syncq, we found it in the list of queues needing
6647  * service, so there is work to be done (or it wouldn't be in the list).
6648  *
6649  * If called from some putnext variation, it was because the
6650  * perimeter is open, but messages are blocking a putnext and
6651  * there is not a thread working on it.  Now a thread could start
6652  * working on it while we are getting ready to do so ourself, but
6653  * the thread would set the q_draining flag, and we can spin out.
6654  *
6655  * As for qwait(_sig), I think I shall let it continue to call
6656  * drain_syncq directly (after all, it will get here eventually).
6657  *
6658  * qdrain_syncq has to terminate when:
6659  * - one of the SQ_STAYAWAY bits gets set to preserve qwriter(OUTER) ordering
6660  * - SQ_EVENTS gets set to preserve qwriter(INNER) ordering
6661  *
6662  * ASSUMES:
6663  *	One claim
6664  *	QLOCK held
6665  *	SQLOCK not held
6666  *	Will release QLOCK before returning
6667  */
6668 void
qdrain_syncq(syncq_t * sq,queue_t * q)6669 qdrain_syncq(syncq_t *sq, queue_t *q)
6670 {
6671 	mblk_t		*bp;
6672 #ifdef DEBUG
6673 	uint16_t	count;
6674 #endif
6675 
6676 	TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_START,
6677 	    "drain_syncq start:%p", sq);
6678 	ASSERT(q->q_syncq == sq);
6679 	ASSERT(MUTEX_HELD(QLOCK(q)));
6680 	ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
6681 	/*
6682 	 * For non-CIPUT perimeters, we should be called with the exclusive bit
6683 	 * set already. For CIPUT perimeters, we will be doing a concurrent
6684 	 * drain, so it better not be set.
6685 	 */
6686 	ASSERT((sq->sq_flags & (SQ_EXCL|SQ_CIPUT)));
6687 	ASSERT(!((sq->sq_type & SQ_CIPUT) && (sq->sq_flags & SQ_EXCL)));
6688 	ASSERT((sq->sq_type & SQ_CIPUT) || (sq->sq_flags & SQ_EXCL));
6689 	/*
6690 	 * All outer pointers are set, or none of them are
6691 	 */
6692 	ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6693 	    sq->sq_oprev == NULL) ||
6694 	    (sq->sq_outer != NULL && sq->sq_onext != NULL &&
6695 	    sq->sq_oprev != NULL));
6696 #ifdef DEBUG
6697 	count = sq->sq_count;
6698 	/*
6699 	 * This is OK without the putlocks, because we have one
6700 	 * claim either from the sq_count, or a putcount.  We could
6701 	 * get an erroneous value from other counts, but ours won't
6702 	 * change, so one way or another, we will have at least a
6703 	 * value of one.
6704 	 */
6705 	SUM_SQ_PUTCOUNTS(sq, count);
6706 	ASSERT(count >= 1);
6707 #endif /* DEBUG */
6708 
6709 	/*
6710 	 * The first thing to do is find out if a thread is already draining
6711 	 * this queue. If so, we are done, just return.
6712 	 */
6713 	if (q->q_draining) {
6714 		mutex_exit(QLOCK(q));
6715 		return;
6716 	}
6717 
6718 	/*
6719 	 * If the perimeter is exclusive, there is nothing we can do right now,
6720 	 * go away. Note that there is nothing to prevent this case from
6721 	 * changing right after this check, but the spin-out will catch it.
6722 	 */
6723 
6724 	/* Tell other threads that we are draining this queue */
6725 	q->q_draining = 1;	/* Protected by QLOCK */
6726 
6727 	/*
6728 	 * If there is nothing to do, clear QFULL as necessary. This caters for
6729 	 * the case where an empty queue was enqueued onto the syncq.
6730 	 */
6731 	if (q->q_sqhead == NULL) {
6732 		ASSERT(q->q_syncqmsgs == 0);
6733 		mutex_exit(QLOCK(q));
6734 		clr_qfull(q);
6735 		mutex_enter(QLOCK(q));
6736 	}
6737 
6738 	/*
6739 	 * Note that q_sqhead must be re-checked here in case another message
6740 	 * was enqueued whilst QLOCK was dropped during the call to clr_qfull.
6741 	 */
6742 	for (bp = q->q_sqhead; bp != NULL; bp = q->q_sqhead) {
6743 		/*
6744 		 * Because we can enter this routine just because a putnext is
6745 		 * blocked, we need to spin out if the perimeter wants to go
6746 		 * exclusive as well as just blocked. We need to spin out also
6747 		 * if events are queued on the syncq.
6748 		 * Don't check for SQ_EXCL, because non-CIPUT perimeters would
6749 		 * set it, and it can't become exclusive while we hold a claim.
6750 		 */
6751 		if (sq->sq_flags & (SQ_STAYAWAY | SQ_EVENTS)) {
6752 			break;
6753 		}
6754 
6755 #ifdef DEBUG
6756 		/*
6757 		 * Since we are in qdrain_syncq, we already know the queue,
6758 		 * but for sanity, we want to check this against the qp that
6759 		 * was passed in by bp->b_queue.
6760 		 */
6761 
6762 		ASSERT(bp->b_queue == q);
6763 		ASSERT(bp->b_queue->q_syncq == sq);
6764 		bp->b_queue = NULL;
6765 
6766 		/*
6767 		 * We would have the following check in the DEBUG code:
6768 		 *
6769 		 * if (bp->b_prev != NULL)  {
6770 		 *	ASSERT(bp->b_prev == (void (*)())q->q_qinfo->qi_putp);
6771 		 * }
6772 		 *
6773 		 * This can't be done, however, since IP modifies qinfo
6774 		 * structure at run-time (switching between IPv4 qinfo and IPv6
6775 		 * qinfo), invalidating the check.
6776 		 * So the assignment to func is left here, but the ASSERT itself
6777 		 * is removed until the whole issue is resolved.
6778 		 */
6779 #endif
6780 		ASSERT(q->q_sqhead == bp);
6781 		q->q_sqhead = bp->b_next;
6782 		bp->b_prev = bp->b_next = NULL;
6783 		ASSERT(q->q_syncqmsgs > 0);
6784 		mutex_exit(QLOCK(q));
6785 
6786 		ASSERT(bp->b_datap->db_ref != 0);
6787 
6788 		(void) (*q->q_qinfo->qi_putp)(q, bp);
6789 
6790 		mutex_enter(QLOCK(q));
6791 
6792 		/*
6793 		 * q_syncqmsgs should only be decremented after executing the
6794 		 * put procedure to avoid message re-ordering. This is due to an
6795 		 * optimisation in putnext() which can call the put procedure
6796 		 * directly if it sees q_syncqmsgs == 0 (despite Q_SQQUEUED
6797 		 * being set).
6798 		 *
6799 		 * We also need to clear QFULL in the next service procedure
6800 		 * queue if this is the last message destined for that queue.
6801 		 *
6802 		 * It would make better sense to have some sort of tunable for
6803 		 * the low water mark, but these semantics are not yet defined.
6804 		 * So, alas, we use a constant.
6805 		 */
6806 		if (--q->q_syncqmsgs == 0) {
6807 			mutex_exit(QLOCK(q));
6808 			clr_qfull(q);
6809 			mutex_enter(QLOCK(q));
6810 		}
6811 
6812 		/*
6813 		 * Always clear SQ_EXCL when CIPUT in order to handle
6814 		 * qwriter(INNER). The putp() can call qwriter and get exclusive
6815 		 * access IFF this is the only claim. So, we need to test for
6816 		 * this possibility, acquire the mutex and clear the bit.
6817 		 */
6818 		if ((sq->sq_type & SQ_CIPUT) && (sq->sq_flags & SQ_EXCL)) {
6819 			mutex_enter(SQLOCK(sq));
6820 			sq->sq_flags &= ~SQ_EXCL;
6821 			mutex_exit(SQLOCK(sq));
6822 		}
6823 	}
6824 
6825 	/*
6826 	 * We should either have no messages on this queue, or we were told to
6827 	 * goaway by a waiter (which we will wake up at the end of this
6828 	 * function).
6829 	 */
6830 	ASSERT((q->q_sqhead == NULL) ||
6831 	    (sq->sq_flags & (SQ_STAYAWAY | SQ_EVENTS)));
6832 
6833 	ASSERT(MUTEX_HELD(QLOCK(q)));
6834 	ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
6835 
6836 	/* Remove the q from the syncq list if all the messages are drained. */
6837 	if (q->q_sqhead == NULL) {
6838 		ASSERT(q->q_syncqmsgs == 0);
6839 		mutex_enter(SQLOCK(sq));
6840 		if (q->q_sqflags & Q_SQQUEUED)
6841 			SQRM_Q(sq, q);
6842 		mutex_exit(SQLOCK(sq));
6843 		/*
6844 		 * Since the queue is removed from the list, reset its priority.
6845 		 */
6846 		q->q_spri = 0;
6847 	}
6848 
6849 	/*
6850 	 * Remember, the q_draining flag is used to let another thread know
6851 	 * that there is a thread currently draining the messages for a queue.
6852 	 * Since we are now done with this queue (even if there may be messages
6853 	 * still there), we need to clear this flag so some thread will work on
6854 	 * it if needed.
6855 	 */
6856 	ASSERT(q->q_draining);
6857 	q->q_draining = 0;
6858 
6859 	/* Called with a claim, so OK to drop all locks. */
6860 	mutex_exit(QLOCK(q));
6861 
6862 	TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_END,
6863 	    "drain_syncq end:%p", sq);
6864 }
6865 /* END OF QDRAIN_SYNCQ  */
6866 
6867 
6868 /*
6869  * This is the mate to qdrain_syncq, except that it is putting the message onto
6870  * the queue instead of draining. Since the message is destined for the queue
6871  * that is selected, there is no need to identify the function because the
6872  * message is intended for the put routine for the queue. For debug kernels,
6873  * this routine will do it anyway just in case.
6874  *
6875  * After the message is enqueued on the syncq, it calls putnext_tail()
6876  * which will schedule a background thread to actually process the message.
6877  *
6878  * Assumes that there is a claim on the syncq (sq->sq_count > 0) and
6879  * SQLOCK(sq) and QLOCK(q) are not held.
6880  */
6881 void
qfill_syncq(syncq_t * sq,queue_t * q,mblk_t * mp)6882 qfill_syncq(syncq_t *sq, queue_t *q, mblk_t *mp)
6883 {
6884 	ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
6885 	ASSERT(MUTEX_NOT_HELD(QLOCK(q)));
6886 	ASSERT(sq->sq_count > 0);
6887 	ASSERT(q->q_syncq == sq);
6888 	ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6889 	    sq->sq_oprev == NULL) ||
6890 	    (sq->sq_outer != NULL && sq->sq_onext != NULL &&
6891 	    sq->sq_oprev != NULL));
6892 
6893 	mutex_enter(QLOCK(q));
6894 
6895 #ifdef DEBUG
6896 	/*
6897 	 * This is used for debug in the qfill_syncq/qdrain_syncq case
6898 	 * to trace the queue that the message is intended for.  Note
6899 	 * that the original use was to identify the queue and function
6900 	 * to call on the drain.  In the new syncq, we have the context
6901 	 * of the queue that we are draining, so call it's putproc and
6902 	 * don't rely on the saved values.  But for debug this is still
6903 	 * useful information.
6904 	 */
6905 	mp->b_prev = (mblk_t *)q->q_qinfo->qi_putp;
6906 	mp->b_queue = q;
6907 	mp->b_next = NULL;
6908 #endif
6909 	ASSERT(q->q_syncq == sq);
6910 	/*
6911 	 * Enqueue the message on the list.
6912 	 * SQPUT_MP() accesses q_syncqmsgs.  We are already holding QLOCK to
6913 	 * protect it.  So it's ok to acquire SQLOCK after SQPUT_MP().
6914 	 */
6915 	SQPUT_MP(q, mp);
6916 	mutex_enter(SQLOCK(sq));
6917 
6918 	/*
6919 	 * And queue on syncq for scheduling, if not already queued.
6920 	 * Note that we need the SQLOCK for this, and for testing flags
6921 	 * at the end to see if we will drain.  So grab it now, and
6922 	 * release it before we call qdrain_syncq or return.
6923 	 */
6924 	if (!(q->q_sqflags & Q_SQQUEUED)) {
6925 		q->q_spri = curthread->t_pri;
6926 		SQPUT_Q(sq, q);
6927 	}
6928 #ifdef DEBUG
6929 	else {
6930 		/*
6931 		 * All of these conditions MUST be true!
6932 		 */
6933 		ASSERT(sq->sq_tail != NULL);
6934 		if (sq->sq_tail == sq->sq_head) {
6935 			ASSERT((q->q_sqprev == NULL) &&
6936 			    (q->q_sqnext == NULL));
6937 		} else {
6938 			ASSERT((q->q_sqprev != NULL) ||
6939 			    (q->q_sqnext != NULL));
6940 		}
6941 		ASSERT(sq->sq_flags & SQ_QUEUED);
6942 		ASSERT(q->q_syncqmsgs != 0);
6943 		ASSERT(q->q_sqflags & Q_SQQUEUED);
6944 	}
6945 #endif
6946 	mutex_exit(QLOCK(q));
6947 	/*
6948 	 * SQLOCK is still held, so sq_count can be safely decremented.
6949 	 */
6950 	sq->sq_count--;
6951 
6952 	putnext_tail(sq, q, 0);
6953 	/* Should not reference sq or q after this point. */
6954 }
6955 
6956 /*  End of qfill_syncq  */
6957 
6958 /*
6959  * Remove all messages from a syncq (if qp is NULL) or remove all messages
6960  * that would be put into qp by drain_syncq.
6961  * Used when deleting the syncq (qp == NULL) or when detaching
6962  * a queue (qp != NULL).
6963  * Return non-zero if one or more messages were freed.
6964  *
6965  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
6966  * sq_putlocks are used.
6967  *
6968  * NOTE: This function assumes that it is called from the close() context and
6969  * that all the queues in the syncq are going away. For this reason it doesn't
6970  * acquire QLOCK for modifying q_sqhead/q_sqtail fields. This assumption is
6971  * currently valid, but it is useful to rethink this function to behave properly
6972  * in other cases.
6973  */
6974 int
flush_syncq(syncq_t * sq,queue_t * qp)6975 flush_syncq(syncq_t *sq, queue_t *qp)
6976 {
6977 	mblk_t		*bp, *mp_head, *mp_next, *mp_prev;
6978 	queue_t		*q;
6979 	int		ret = 0;
6980 
6981 	mutex_enter(SQLOCK(sq));
6982 
6983 	/*
6984 	 * Before we leave, we need to make sure there are no
6985 	 * events listed for this queue.  All events for this queue
6986 	 * will just be freed.
6987 	 */
6988 	if (qp != NULL && sq->sq_evhead != NULL) {
6989 		ASSERT(sq->sq_flags & SQ_EVENTS);
6990 
6991 		mp_prev = NULL;
6992 		for (bp = sq->sq_evhead; bp != NULL; bp = mp_next) {
6993 			mp_next = bp->b_next;
6994 			if (bp->b_queue == qp) {
6995 				/* Delete this message */
6996 				if (mp_prev != NULL) {
6997 					mp_prev->b_next = mp_next;
6998 					/*
6999 					 * Update sq_evtail if the last element
7000 					 * is removed.
7001 					 */
7002 					if (bp == sq->sq_evtail) {
7003 						ASSERT(mp_next == NULL);
7004 						sq->sq_evtail = mp_prev;
7005 					}
7006 				} else
7007 					sq->sq_evhead = mp_next;
7008 				if (sq->sq_evhead == NULL)
7009 					sq->sq_flags &= ~SQ_EVENTS;
7010 				bp->b_prev = bp->b_next = NULL;
7011 				freemsg(bp);
7012 				ret++;
7013 			} else {
7014 				mp_prev = bp;
7015 			}
7016 		}
7017 	}
7018 
7019 	/*
7020 	 * Walk sq_head and:
7021 	 *	- match qp if qp is set, remove it's messages
7022 	 *	- all if qp is not set
7023 	 */
7024 	q = sq->sq_head;
7025 	while (q != NULL) {
7026 		ASSERT(q->q_syncq == sq);
7027 		if ((qp == NULL) || (qp == q)) {
7028 			/*
7029 			 * Yank the messages as a list off the queue
7030 			 */
7031 			mp_head = q->q_sqhead;
7032 			/*
7033 			 * We do not have QLOCK(q) here (which is safe due to
7034 			 * assumptions mentioned above). To obtain the lock we
7035 			 * need to release SQLOCK which may allow lots of things
7036 			 * to change upon us. This place requires more analysis.
7037 			 */
7038 			q->q_sqhead = q->q_sqtail = NULL;
7039 			ASSERT(mp_head->b_queue &&
7040 			    mp_head->b_queue->q_syncq == sq);
7041 
7042 			/*
7043 			 * Free each of the messages.
7044 			 */
7045 			for (bp = mp_head; bp != NULL; bp = mp_next) {
7046 				mp_next = bp->b_next;
7047 				bp->b_prev = bp->b_next = NULL;
7048 				freemsg(bp);
7049 				ret++;
7050 			}
7051 			/*
7052 			 * Now remove the queue from the syncq.
7053 			 */
7054 			ASSERT(q->q_sqflags & Q_SQQUEUED);
7055 			SQRM_Q(sq, q);
7056 			q->q_spri = 0;
7057 			q->q_syncqmsgs = 0;
7058 
7059 			/*
7060 			 * If qp was specified, we are done with it and are
7061 			 * going to drop SQLOCK(sq) and return. We wakeup syncq
7062 			 * waiters while we still have the SQLOCK.
7063 			 */
7064 			if ((qp != NULL) && (sq->sq_flags & SQ_WANTWAKEUP)) {
7065 				sq->sq_flags &= ~SQ_WANTWAKEUP;
7066 				cv_broadcast(&sq->sq_wait);
7067 			}
7068 			/* Drop SQLOCK across clr_qfull */
7069 			mutex_exit(SQLOCK(sq));
7070 
7071 			/*
7072 			 * We avoid doing the test that drain_syncq does and
7073 			 * unconditionally clear qfull for every flushed
7074 			 * message. Since flush_syncq is only called during
7075 			 * close this should not be a problem.
7076 			 */
7077 			clr_qfull(q);
7078 			if (qp != NULL) {
7079 				return (ret);
7080 			} else {
7081 				mutex_enter(SQLOCK(sq));
7082 				/*
7083 				 * The head was removed by SQRM_Q above.
7084 				 * reread the new head and flush it.
7085 				 */
7086 				q = sq->sq_head;
7087 			}
7088 		} else {
7089 			q = q->q_sqnext;
7090 		}
7091 		ASSERT(MUTEX_HELD(SQLOCK(sq)));
7092 	}
7093 
7094 	if (sq->sq_flags & SQ_WANTWAKEUP) {
7095 		sq->sq_flags &= ~SQ_WANTWAKEUP;
7096 		cv_broadcast(&sq->sq_wait);
7097 	}
7098 
7099 	mutex_exit(SQLOCK(sq));
7100 	return (ret);
7101 }
7102 
7103 /*
7104  * Propagate all messages from a syncq to the next syncq that are associated
7105  * with the specified queue. If the queue is attached to a driver or if the
7106  * messages have been added due to a qwriter(PERIM_INNER), free the messages.
7107  *
7108  * Assumes that the stream is strlock()'ed. We don't come here if there
7109  * are no messages to propagate.
7110  *
7111  * NOTE : If the queue is attached to a driver, all the messages are freed
7112  * as there is no point in propagating the messages from the driver syncq
7113  * to the closing stream head which will in turn get freed later.
7114  */
7115 static int
propagate_syncq(queue_t * qp)7116 propagate_syncq(queue_t *qp)
7117 {
7118 	mblk_t		*bp, *head, *tail, *prev, *next;
7119 	syncq_t		*sq;
7120 	queue_t		*nqp;
7121 	syncq_t		*nsq;
7122 	boolean_t	isdriver;
7123 	int		moved = 0;
7124 	uint16_t	flags;
7125 	pri_t		priority = curthread->t_pri;
7126 #ifdef DEBUG
7127 	void		(*func)();
7128 #endif
7129 
7130 	sq = qp->q_syncq;
7131 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
7132 	/* debug macro */
7133 	SQ_PUTLOCKS_HELD(sq);
7134 	/*
7135 	 * As entersq() does not increment the sq_count for
7136 	 * the write side, check sq_count for non-QPERQ
7137 	 * perimeters alone.
7138 	 */
7139 	ASSERT((qp->q_flag & QPERQ) || (sq->sq_count >= 1));
7140 
7141 	/*
7142 	 * propagate_syncq() can be called because of either messages on the
7143 	 * queue syncq or because on events on the queue syncq. Do actual
7144 	 * message propagations if there are any messages.
7145 	 */
7146 	if (qp->q_syncqmsgs) {
7147 		isdriver = (qp->q_flag & QISDRV);
7148 
7149 		if (!isdriver) {
7150 			nqp = qp->q_next;
7151 			nsq = nqp->q_syncq;
7152 			ASSERT(MUTEX_HELD(SQLOCK(nsq)));
7153 			/* debug macro */
7154 			SQ_PUTLOCKS_HELD(nsq);
7155 #ifdef DEBUG
7156 			func = (void (*)())(uintptr_t)nqp->q_qinfo->qi_putp;
7157 #endif
7158 		}
7159 
7160 		SQRM_Q(sq, qp);
7161 		priority = MAX(qp->q_spri, priority);
7162 		qp->q_spri = 0;
7163 		head = qp->q_sqhead;
7164 		tail = qp->q_sqtail;
7165 		qp->q_sqhead = qp->q_sqtail = NULL;
7166 		qp->q_syncqmsgs = 0;
7167 
7168 		/*
7169 		 * Walk the list of messages, and free them if this is a driver,
7170 		 * otherwise reset the b_prev and b_queue value to the new putp.
7171 		 * Afterward, we will just add the head to the end of the next
7172 		 * syncq, and point the tail to the end of this one.
7173 		 */
7174 
7175 		for (bp = head; bp != NULL; bp = next) {
7176 			next = bp->b_next;
7177 			if (isdriver) {
7178 				bp->b_prev = bp->b_next = NULL;
7179 				freemsg(bp);
7180 				continue;
7181 			}
7182 			/* Change the q values for this message */
7183 			bp->b_queue = nqp;
7184 #ifdef DEBUG
7185 			bp->b_prev = (mblk_t *)func;
7186 #endif
7187 			moved++;
7188 		}
7189 		/*
7190 		 * Attach list of messages to the end of the new queue (if there
7191 		 * is a list of messages).
7192 		 */
7193 
7194 		if (!isdriver && head != NULL) {
7195 			ASSERT(tail != NULL);
7196 			if (nqp->q_sqhead == NULL) {
7197 				nqp->q_sqhead = head;
7198 			} else {
7199 				ASSERT(nqp->q_sqtail != NULL);
7200 				nqp->q_sqtail->b_next = head;
7201 			}
7202 			nqp->q_sqtail = tail;
7203 			/*
7204 			 * When messages are moved from high priority queue to
7205 			 * another queue, the destination queue priority is
7206 			 * upgraded.
7207 			 */
7208 
7209 			if (priority > nqp->q_spri)
7210 				nqp->q_spri = priority;
7211 
7212 			SQPUT_Q(nsq, nqp);
7213 
7214 			nqp->q_syncqmsgs += moved;
7215 			ASSERT(nqp->q_syncqmsgs != 0);
7216 		}
7217 	}
7218 
7219 	/*
7220 	 * Before we leave, we need to make sure there are no
7221 	 * events listed for this queue.  All events for this queue
7222 	 * will just be freed.
7223 	 */
7224 	if (sq->sq_evhead != NULL) {
7225 		ASSERT(sq->sq_flags & SQ_EVENTS);
7226 		prev = NULL;
7227 		for (bp = sq->sq_evhead; bp != NULL; bp = next) {
7228 			next = bp->b_next;
7229 			if (bp->b_queue == qp) {
7230 				/* Delete this message */
7231 				if (prev != NULL) {
7232 					prev->b_next = next;
7233 					/*
7234 					 * Update sq_evtail if the last element
7235 					 * is removed.
7236 					 */
7237 					if (bp == sq->sq_evtail) {
7238 						ASSERT(next == NULL);
7239 						sq->sq_evtail = prev;
7240 					}
7241 				} else
7242 					sq->sq_evhead = next;
7243 				if (sq->sq_evhead == NULL)
7244 					sq->sq_flags &= ~SQ_EVENTS;
7245 				bp->b_prev = bp->b_next = NULL;
7246 				freemsg(bp);
7247 			} else {
7248 				prev = bp;
7249 			}
7250 		}
7251 	}
7252 
7253 	flags = sq->sq_flags;
7254 
7255 	/* Wake up any waiter before leaving. */
7256 	if (flags & SQ_WANTWAKEUP) {
7257 		flags &= ~SQ_WANTWAKEUP;
7258 		cv_broadcast(&sq->sq_wait);
7259 	}
7260 	sq->sq_flags = flags;
7261 
7262 	return (moved);
7263 }
7264 
7265 /*
7266  * Try and upgrade to exclusive access at the inner perimeter. If this can
7267  * not be done without blocking then request will be queued on the syncq
7268  * and drain_syncq will run it later.
7269  *
7270  * This routine can only be called from put or service procedures plus
7271  * asynchronous callback routines that have properly entered the queue (with
7272  * entersq). Thus qwriter_inner assumes the caller has one claim on the syncq
7273  * associated with q.
7274  */
7275 void
qwriter_inner(queue_t * q,mblk_t * mp,void (* func)())7276 qwriter_inner(queue_t *q, mblk_t *mp, void (*func)())
7277 {
7278 	syncq_t	*sq = q->q_syncq;
7279 	uint16_t count;
7280 
7281 	mutex_enter(SQLOCK(sq));
7282 	count = sq->sq_count;
7283 	SQ_PUTLOCKS_ENTER(sq);
7284 	SUM_SQ_PUTCOUNTS(sq, count);
7285 	ASSERT(count >= 1);
7286 	ASSERT(sq->sq_type & (SQ_CIPUT|SQ_CISVC));
7287 
7288 	if (count == 1) {
7289 		/*
7290 		 * Can upgrade. This case also handles nested qwriter calls
7291 		 * (when the qwriter callback function calls qwriter). In that
7292 		 * case SQ_EXCL is already set.
7293 		 */
7294 		sq->sq_flags |= SQ_EXCL;
7295 		SQ_PUTLOCKS_EXIT(sq);
7296 		mutex_exit(SQLOCK(sq));
7297 		(*func)(q, mp);
7298 		/*
7299 		 * Assumes that leavesq, putnext, and drain_syncq will reset
7300 		 * SQ_EXCL for SQ_CIPUT/SQ_CISVC queues. We leave SQ_EXCL on
7301 		 * until putnext, leavesq, or drain_syncq drops it.
7302 		 * That way we handle nested qwriter(INNER) without dropping
7303 		 * SQ_EXCL until the outermost qwriter callback routine is
7304 		 * done.
7305 		 */
7306 		return;
7307 	}
7308 	SQ_PUTLOCKS_EXIT(sq);
7309 	sqfill_events(sq, q, mp, func);
7310 }
7311 
7312 /*
7313  * Synchronous callback support functions
7314  */
7315 
7316 /*
7317  * Allocate a callback parameter structure.
7318  * Assumes that caller initializes the flags and the id.
7319  * Acquires SQLOCK(sq) if non-NULL is returned.
7320  */
7321 callbparams_t *
callbparams_alloc(syncq_t * sq,void (* func)(void *),void * arg,int kmflags)7322 callbparams_alloc(syncq_t *sq, void (*func)(void *), void *arg, int kmflags)
7323 {
7324 	callbparams_t *cbp;
7325 	size_t size = sizeof (callbparams_t);
7326 
7327 	cbp = kmem_alloc(size, kmflags & ~KM_PANIC);
7328 
7329 	/*
7330 	 * Only try tryhard allocation if the caller is ready to panic.
7331 	 * Otherwise just fail.
7332 	 */
7333 	if (cbp == NULL) {
7334 		if (kmflags & KM_PANIC)
7335 			cbp = kmem_alloc_tryhard(sizeof (callbparams_t),
7336 			    &size, kmflags);
7337 		else
7338 			return (NULL);
7339 	}
7340 
7341 	ASSERT(size >= sizeof (callbparams_t));
7342 	cbp->cbp_size = size;
7343 	cbp->cbp_sq = sq;
7344 	cbp->cbp_func = func;
7345 	cbp->cbp_arg = arg;
7346 	mutex_enter(SQLOCK(sq));
7347 	cbp->cbp_next = sq->sq_callbpend;
7348 	sq->sq_callbpend = cbp;
7349 	return (cbp);
7350 }
7351 
7352 void
callbparams_free(syncq_t * sq,callbparams_t * cbp)7353 callbparams_free(syncq_t *sq, callbparams_t *cbp)
7354 {
7355 	callbparams_t **pp, *p;
7356 
7357 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
7358 
7359 	for (pp = &sq->sq_callbpend; (p = *pp) != NULL; pp = &p->cbp_next) {
7360 		if (p == cbp) {
7361 			*pp = p->cbp_next;
7362 			kmem_free(p, p->cbp_size);
7363 			return;
7364 		}
7365 	}
7366 	(void) (STRLOG(0, 0, 0, SL_CONSOLE,
7367 	    "callbparams_free: not found\n"));
7368 }
7369 
7370 void
callbparams_free_id(syncq_t * sq,callbparams_id_t id,int32_t flag)7371 callbparams_free_id(syncq_t *sq, callbparams_id_t id, int32_t flag)
7372 {
7373 	callbparams_t **pp, *p;
7374 
7375 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
7376 
7377 	for (pp = &sq->sq_callbpend; (p = *pp) != NULL; pp = &p->cbp_next) {
7378 		if (p->cbp_id == id && p->cbp_flags == flag) {
7379 			*pp = p->cbp_next;
7380 			kmem_free(p, p->cbp_size);
7381 			return;
7382 		}
7383 	}
7384 	(void) (STRLOG(0, 0, 0, SL_CONSOLE,
7385 	    "callbparams_free_id: not found\n"));
7386 }
7387 
7388 /*
7389  * Callback wrapper function used by once-only callbacks that can be
7390  * cancelled (qtimeout and qbufcall)
7391  * Contains inline version of entersq(sq, SQ_CALLBACK) that can be
7392  * cancelled by the qun* functions.
7393  */
7394 void
qcallbwrapper(void * arg)7395 qcallbwrapper(void *arg)
7396 {
7397 	callbparams_t *cbp = arg;
7398 	syncq_t	*sq;
7399 	uint16_t count = 0;
7400 	uint16_t waitflags = SQ_STAYAWAY | SQ_EVENTS | SQ_EXCL;
7401 	uint16_t type;
7402 
7403 	sq = cbp->cbp_sq;
7404 	mutex_enter(SQLOCK(sq));
7405 	type = sq->sq_type;
7406 	if (!(type & SQ_CICB)) {
7407 		count = sq->sq_count;
7408 		SQ_PUTLOCKS_ENTER(sq);
7409 		SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
7410 		SUM_SQ_PUTCOUNTS(sq, count);
7411 		sq->sq_needexcl++;
7412 		ASSERT(sq->sq_needexcl != 0);	/* wraparound */
7413 		waitflags |= SQ_MESSAGES;
7414 	}
7415 	/* Can not handle exclusive entry at outer perimeter */
7416 	ASSERT(type & SQ_COCB);
7417 
7418 	while ((sq->sq_flags & waitflags) || (!(type & SQ_CICB) &&count != 0)) {
7419 		if ((sq->sq_callbflags & cbp->cbp_flags) &&
7420 		    (sq->sq_cancelid == cbp->cbp_id)) {
7421 			/* timeout has been cancelled */
7422 			sq->sq_callbflags |= SQ_CALLB_BYPASSED;
7423 			callbparams_free(sq, cbp);
7424 			if (!(type & SQ_CICB)) {
7425 				ASSERT(sq->sq_needexcl > 0);
7426 				sq->sq_needexcl--;
7427 				if (sq->sq_needexcl == 0) {
7428 					SQ_PUTCOUNT_SETFAST_LOCKED(sq);
7429 				}
7430 				SQ_PUTLOCKS_EXIT(sq);
7431 			}
7432 			mutex_exit(SQLOCK(sq));
7433 			return;
7434 		}
7435 		sq->sq_flags |= SQ_WANTWAKEUP;
7436 		if (!(type & SQ_CICB)) {
7437 			SQ_PUTLOCKS_EXIT(sq);
7438 		}
7439 		cv_wait(&sq->sq_wait, SQLOCK(sq));
7440 		if (!(type & SQ_CICB)) {
7441 			count = sq->sq_count;
7442 			SQ_PUTLOCKS_ENTER(sq);
7443 			SUM_SQ_PUTCOUNTS(sq, count);
7444 		}
7445 	}
7446 
7447 	sq->sq_count++;
7448 	ASSERT(sq->sq_count != 0);	/* Wraparound */
7449 	if (!(type & SQ_CICB)) {
7450 		ASSERT(count == 0);
7451 		sq->sq_flags |= SQ_EXCL;
7452 		ASSERT(sq->sq_needexcl > 0);
7453 		sq->sq_needexcl--;
7454 		if (sq->sq_needexcl == 0) {
7455 			SQ_PUTCOUNT_SETFAST_LOCKED(sq);
7456 		}
7457 		SQ_PUTLOCKS_EXIT(sq);
7458 	}
7459 
7460 	mutex_exit(SQLOCK(sq));
7461 
7462 	cbp->cbp_func(cbp->cbp_arg);
7463 
7464 	/*
7465 	 * We drop the lock only for leavesq to re-acquire it.
7466 	 * Possible optimization is inline of leavesq.
7467 	 */
7468 	mutex_enter(SQLOCK(sq));
7469 	callbparams_free(sq, cbp);
7470 	mutex_exit(SQLOCK(sq));
7471 	leavesq(sq, SQ_CALLBACK);
7472 }
7473 
7474 /*
7475  * No need to grab sq_putlocks here. See comment in strsubr.h that
7476  * explains when sq_putlocks are used.
7477  *
7478  * sq_count (or one of the sq_putcounts) has already been
7479  * decremented by the caller, and if SQ_QUEUED, we need to call
7480  * drain_syncq (the global syncq drain).
7481  * If putnext_tail is called with the SQ_EXCL bit set, we are in
7482  * one of two states, non-CIPUT perimeter, and we need to clear
7483  * it, or we went exclusive in the put procedure.  In any case,
7484  * we want to clear the bit now, and it is probably easier to do
7485  * this at the beginning of this function (remember, we hold
7486  * the SQLOCK).  Lastly, if there are other messages queued
7487  * on the syncq (and not for our destination), enable the syncq
7488  * for background work.
7489  */
7490 
7491 /* ARGSUSED */
7492 void
putnext_tail(syncq_t * sq,queue_t * qp,uint32_t passflags)7493 putnext_tail(syncq_t *sq, queue_t *qp, uint32_t passflags)
7494 {
7495 	uint16_t	flags = sq->sq_flags;
7496 
7497 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
7498 	ASSERT(MUTEX_NOT_HELD(QLOCK(qp)));
7499 
7500 	/* Clear SQ_EXCL if set in passflags */
7501 	if (passflags & SQ_EXCL) {
7502 		flags &= ~SQ_EXCL;
7503 	}
7504 	if (flags & SQ_WANTWAKEUP) {
7505 		flags &= ~SQ_WANTWAKEUP;
7506 		cv_broadcast(&sq->sq_wait);
7507 	}
7508 	if (flags & SQ_WANTEXWAKEUP) {
7509 		flags &= ~SQ_WANTEXWAKEUP;
7510 		cv_broadcast(&sq->sq_exitwait);
7511 	}
7512 	sq->sq_flags = flags;
7513 
7514 	/*
7515 	 * We have cleared SQ_EXCL if we were asked to, and started
7516 	 * the wakeup process for waiters.  If there are no writers
7517 	 * then we need to drain the syncq if we were told to, or
7518 	 * enable the background thread to do it.
7519 	 */
7520 	if (!(flags & (SQ_STAYAWAY|SQ_EXCL))) {
7521 		if ((passflags & SQ_QUEUED) ||
7522 		    (sq->sq_svcflags & SQ_DISABLED)) {
7523 			/* drain_syncq will take care of events in the list */
7524 			drain_syncq(sq);
7525 			return;
7526 		} else if (flags & SQ_QUEUED) {
7527 			sqenable(sq);
7528 		}
7529 	}
7530 	/* Drop the SQLOCK on exit */
7531 	mutex_exit(SQLOCK(sq));
7532 	TRACE_3(TR_FAC_STREAMS_FR, TR_PUTNEXT_END,
7533 	    "putnext_end:(%p, %p, %p) done", NULL, qp, sq);
7534 }
7535 
7536 void
set_qend(queue_t * q)7537 set_qend(queue_t *q)
7538 {
7539 	mutex_enter(QLOCK(q));
7540 	if (!O_SAMESTR(q))
7541 		q->q_flag |= QEND;
7542 	else
7543 		q->q_flag &= ~QEND;
7544 	mutex_exit(QLOCK(q));
7545 	q = _OTHERQ(q);
7546 	mutex_enter(QLOCK(q));
7547 	if (!O_SAMESTR(q))
7548 		q->q_flag |= QEND;
7549 	else
7550 		q->q_flag &= ~QEND;
7551 	mutex_exit(QLOCK(q));
7552 }
7553 
7554 /*
7555  * Set QFULL in next service procedure queue (that cares) if not already
7556  * set and if there are already more messages on the syncq than
7557  * sq_max_size.  If sq_max_size is 0, no flow control will be asserted on
7558  * any syncq.
7559  *
7560  * The fq here is the next queue with a service procedure.  This is where
7561  * we would fail canputnext, so this is where we need to set QFULL.
7562  * In the case when fq != q we need to take QLOCK(fq) to set QFULL flag.
7563  *
7564  * We already have QLOCK at this point. To avoid cross-locks with
7565  * freezestr() which grabs all QLOCKs and with strlock() which grabs both
7566  * SQLOCK and sd_reflock, we need to drop respective locks first.
7567  */
7568 void
set_qfull(queue_t * q)7569 set_qfull(queue_t *q)
7570 {
7571 	queue_t		*fq = NULL;
7572 
7573 	ASSERT(MUTEX_HELD(QLOCK(q)));
7574 	if ((sq_max_size != 0) && (!(q->q_nfsrv->q_flag & QFULL)) &&
7575 	    (q->q_syncqmsgs > sq_max_size)) {
7576 		if ((fq = q->q_nfsrv) == q) {
7577 			fq->q_flag |= QFULL;
7578 		} else {
7579 			mutex_exit(QLOCK(q));
7580 			mutex_enter(QLOCK(fq));
7581 			fq->q_flag |= QFULL;
7582 			mutex_exit(QLOCK(fq));
7583 			mutex_enter(QLOCK(q));
7584 		}
7585 	}
7586 }
7587 
7588 void
clr_qfull(queue_t * q)7589 clr_qfull(queue_t *q)
7590 {
7591 	queue_t	*oq = q;
7592 
7593 	q = q->q_nfsrv;
7594 	/* Fast check if there is any work to do before getting the lock. */
7595 	if ((q->q_flag & (QFULL|QWANTW)) == 0) {
7596 		return;
7597 	}
7598 
7599 	/*
7600 	 * Do not reset QFULL (and backenable) if the q_count is the reason
7601 	 * for QFULL being set.
7602 	 */
7603 	mutex_enter(QLOCK(q));
7604 	/*
7605 	 * If queue is empty i.e q_mblkcnt is zero, queue can not be full.
7606 	 * Hence clear the QFULL.
7607 	 * If both q_count and q_mblkcnt are less than the hiwat mark,
7608 	 * clear the QFULL.
7609 	 */
7610 	if (q->q_mblkcnt == 0 || ((q->q_count < q->q_hiwat) &&
7611 	    (q->q_mblkcnt < q->q_hiwat))) {
7612 		q->q_flag &= ~QFULL;
7613 		/*
7614 		 * A little more confusing, how about this way:
7615 		 * if someone wants to write,
7616 		 * AND
7617 		 *    both counts are less than the lowat mark
7618 		 *    OR
7619 		 *    the lowat mark is zero
7620 		 * THEN
7621 		 * backenable
7622 		 */
7623 		if ((q->q_flag & QWANTW) &&
7624 		    (((q->q_count < q->q_lowat) &&
7625 		    (q->q_mblkcnt < q->q_lowat)) || q->q_lowat == 0)) {
7626 			q->q_flag &= ~QWANTW;
7627 			mutex_exit(QLOCK(q));
7628 			backenable(oq, 0);
7629 		} else
7630 			mutex_exit(QLOCK(q));
7631 	} else
7632 		mutex_exit(QLOCK(q));
7633 }
7634 
7635 /*
7636  * Set the forward service procedure pointer.
7637  *
7638  * Called at insert-time to cache a queue's next forward service procedure in
7639  * q_nfsrv; used by canput() and canputnext().  If the queue to be inserted
7640  * has a service procedure then q_nfsrv points to itself.  If the queue to be
7641  * inserted does not have a service procedure, then q_nfsrv points to the next
7642  * queue forward that has a service procedure.  If the queue is at the logical
7643  * end of the stream (driver for write side, stream head for the read side)
7644  * and does not have a service procedure, then q_nfsrv also points to itself.
7645  */
7646 void
set_nfsrv_ptr(queue_t * rnew,queue_t * wnew,queue_t * prev_rq,queue_t * prev_wq)7647 set_nfsrv_ptr(
7648 	queue_t  *rnew,		/* read queue pointer to new module */
7649 	queue_t  *wnew,		/* write queue pointer to new module */
7650 	queue_t  *prev_rq,	/* read queue pointer to the module above */
7651 	queue_t  *prev_wq)	/* write queue pointer to the module above */
7652 {
7653 	queue_t *qp;
7654 
7655 	if (prev_wq->q_next == NULL) {
7656 		/*
7657 		 * Insert the driver, initialize the driver and stream head.
7658 		 * In this case, prev_rq/prev_wq should be the stream head.
7659 		 * _I_INSERT does not allow inserting a driver.  Make sure
7660 		 * that it is not an insertion.
7661 		 */
7662 		ASSERT(!(rnew->q_flag & _QINSERTING));
7663 		wnew->q_nfsrv = wnew;
7664 		if (rnew->q_qinfo->qi_srvp)
7665 			rnew->q_nfsrv = rnew;
7666 		else
7667 			rnew->q_nfsrv = prev_rq;
7668 		prev_rq->q_nfsrv = prev_rq;
7669 		prev_wq->q_nfsrv = prev_wq;
7670 	} else {
7671 		/*
7672 		 * set up read side q_nfsrv pointer.  This MUST be done
7673 		 * before setting the write side, because the setting of
7674 		 * the write side for a fifo may depend on it.
7675 		 *
7676 		 * Suppose we have a fifo that only has pipemod pushed.
7677 		 * pipemod has no read or write service procedures, so
7678 		 * nfsrv for both pipemod queues points to prev_rq (the
7679 		 * stream read head).  Now push bufmod (which has only a
7680 		 * read service procedure).  Doing the write side first,
7681 		 * wnew->q_nfsrv is set to pipemod's writeq nfsrv, which
7682 		 * is WRONG; the next queue forward from wnew with a
7683 		 * service procedure will be rnew, not the stream read head.
7684 		 * Since the downstream queue (which in the case of a fifo
7685 		 * is the read queue rnew) can affect upstream queues, it
7686 		 * needs to be done first.  Setting up the read side first
7687 		 * sets nfsrv for both pipemod queues to rnew and then
7688 		 * when the write side is set up, wnew-q_nfsrv will also
7689 		 * point to rnew.
7690 		 */
7691 		if (rnew->q_qinfo->qi_srvp) {
7692 			/*
7693 			 * use _OTHERQ() because, if this is a pipe, next
7694 			 * module may have been pushed from other end and
7695 			 * q_next could be a read queue.
7696 			 */
7697 			qp = _OTHERQ(prev_wq->q_next);
7698 			while (qp && qp->q_nfsrv != qp) {
7699 				qp->q_nfsrv = rnew;
7700 				qp = backq(qp);
7701 			}
7702 			rnew->q_nfsrv = rnew;
7703 		} else
7704 			rnew->q_nfsrv = prev_rq->q_nfsrv;
7705 
7706 		/* set up write side q_nfsrv pointer */
7707 		if (wnew->q_qinfo->qi_srvp) {
7708 			wnew->q_nfsrv = wnew;
7709 
7710 			/*
7711 			 * For insertion, need to update nfsrv of the modules
7712 			 * above which do not have a service routine.
7713 			 */
7714 			if (rnew->q_flag & _QINSERTING) {
7715 				for (qp = prev_wq;
7716 				    qp != NULL && qp->q_nfsrv != qp;
7717 				    qp = backq(qp)) {
7718 					qp->q_nfsrv = wnew->q_nfsrv;
7719 				}
7720 			}
7721 		} else {
7722 			if (prev_wq->q_next == prev_rq)
7723 				/*
7724 				 * Since prev_wq/prev_rq are the middle of a
7725 				 * fifo, wnew/rnew will also be the middle of
7726 				 * a fifo and wnew's nfsrv is same as rnew's.
7727 				 */
7728 				wnew->q_nfsrv = rnew->q_nfsrv;
7729 			else
7730 				wnew->q_nfsrv = prev_wq->q_next->q_nfsrv;
7731 		}
7732 	}
7733 }
7734 
7735 /*
7736  * Reset the forward service procedure pointer; called at remove-time.
7737  */
7738 void
reset_nfsrv_ptr(queue_t * rqp,queue_t * wqp)7739 reset_nfsrv_ptr(queue_t *rqp, queue_t *wqp)
7740 {
7741 	queue_t *tmp_qp;
7742 
7743 	/* Reset the write side q_nfsrv pointer for _I_REMOVE */
7744 	if ((rqp->q_flag & _QREMOVING) && (wqp->q_qinfo->qi_srvp != NULL)) {
7745 		for (tmp_qp = backq(wqp);
7746 		    tmp_qp != NULL && tmp_qp->q_nfsrv == wqp;
7747 		    tmp_qp = backq(tmp_qp)) {
7748 			tmp_qp->q_nfsrv = wqp->q_nfsrv;
7749 		}
7750 	}
7751 
7752 	/* reset the read side q_nfsrv pointer */
7753 	if (rqp->q_qinfo->qi_srvp) {
7754 		if (wqp->q_next) {	/* non-driver case */
7755 			tmp_qp = _OTHERQ(wqp->q_next);
7756 			while (tmp_qp && tmp_qp->q_nfsrv == rqp) {
7757 				/* Note that rqp->q_next cannot be NULL */
7758 				ASSERT(rqp->q_next != NULL);
7759 				tmp_qp->q_nfsrv = rqp->q_next->q_nfsrv;
7760 				tmp_qp = backq(tmp_qp);
7761 			}
7762 		}
7763 	}
7764 }
7765 
7766 /*
7767  * This routine should be called after all stream geometry changes to update
7768  * the stream head cached struio() rd/wr queue pointers. Note must be called
7769  * with the streamlock()ed.
7770  *
7771  * Note: only enables Synchronous STREAMS for a side of a Stream which has
7772  *	 an explicit synchronous barrier module queue. That is, a queue that
7773  *	 has specified a struio() type.
7774  */
7775 static void
strsetuio(stdata_t * stp)7776 strsetuio(stdata_t *stp)
7777 {
7778 	queue_t *wrq;
7779 
7780 	if (stp->sd_flag & STPLEX) {
7781 		/*
7782 		 * Not streamhead, but a mux, so no Synchronous STREAMS.
7783 		 */
7784 		stp->sd_struiowrq = NULL;
7785 		stp->sd_struiordq = NULL;
7786 		return;
7787 	}
7788 	/*
7789 	 * Scan the write queue(s) while synchronous
7790 	 * until we find a qinfo uio type specified.
7791 	 */
7792 	wrq = stp->sd_wrq->q_next;
7793 	while (wrq) {
7794 		if (wrq->q_struiot == STRUIOT_NONE) {
7795 			wrq = 0;
7796 			break;
7797 		}
7798 		if (wrq->q_struiot != STRUIOT_DONTCARE)
7799 			break;
7800 		if (! _SAMESTR(wrq)) {
7801 			wrq = 0;
7802 			break;
7803 		}
7804 		wrq = wrq->q_next;
7805 	}
7806 	stp->sd_struiowrq = wrq;
7807 	/*
7808 	 * Scan the read queue(s) while synchronous
7809 	 * until we find a qinfo uio type specified.
7810 	 */
7811 	wrq = stp->sd_wrq->q_next;
7812 	while (wrq) {
7813 		if (_RD(wrq)->q_struiot == STRUIOT_NONE) {
7814 			wrq = 0;
7815 			break;
7816 		}
7817 		if (_RD(wrq)->q_struiot != STRUIOT_DONTCARE)
7818 			break;
7819 		if (! _SAMESTR(wrq)) {
7820 			wrq = 0;
7821 			break;
7822 		}
7823 		wrq = wrq->q_next;
7824 	}
7825 	stp->sd_struiordq = wrq ? _RD(wrq) : 0;
7826 }
7827 
7828 static int
pass_rput(queue_t * q,mblk_t * mp)7829 pass_rput(queue_t *q, mblk_t *mp)
7830 {
7831 	putnext(q, mp);
7832 	return (0);
7833 }
7834 
7835 /*
7836  * pass_wput, unblocks the passthru queues, so that
7837  * messages can arrive at muxs lower read queue, before
7838  * I_LINK/I_UNLINK is acked/nacked.
7839  */
7840 static int
pass_wput(queue_t * q,mblk_t * mp)7841 pass_wput(queue_t *q, mblk_t *mp)
7842 {
7843 	syncq_t *sq;
7844 
7845 	sq = _RD(q)->q_syncq;
7846 	if (sq->sq_flags & SQ_BLOCKED)
7847 		unblocksq(sq, SQ_BLOCKED, 0);
7848 	putnext(q, mp);
7849 	return (0);
7850 }
7851 
7852 /*
7853  * Set up queues for the link/unlink.
7854  * Create a new queue and block it and then insert it
7855  * below the stream head on the lower stream.
7856  * This prevents any messages from arriving during the setq
7857  * as well as while the mux is processing the LINK/I_UNLINK.
7858  * The blocked passq is unblocked once the LINK/I_UNLINK has
7859  * been acked or nacked or if a message is generated and sent
7860  * down muxs write put procedure.
7861  * See pass_wput().
7862  *
7863  * After the new queue is inserted, all messages coming from below are
7864  * blocked. The call to strlock will ensure that all activity in the stream head
7865  * read queue syncq is stopped (sq_count drops to zero).
7866  */
7867 static queue_t *
link_addpassthru(stdata_t * stpdown)7868 link_addpassthru(stdata_t *stpdown)
7869 {
7870 	queue_t *passq;
7871 	sqlist_t sqlist;
7872 
7873 	passq = allocq();
7874 	STREAM(passq) = STREAM(_WR(passq)) = stpdown;
7875 	/* setq might sleep in allocator - avoid holding locks. */
7876 	setq(passq, &passthru_rinit, &passthru_winit, NULL, QPERQ,
7877 	    SQ_CI|SQ_CO, B_FALSE);
7878 	claimq(passq);
7879 	blocksq(passq->q_syncq, SQ_BLOCKED, 1);
7880 	insertq(STREAM(passq), passq);
7881 
7882 	/*
7883 	 * Use strlock() to wait for the stream head sq_count to drop to zero
7884 	 * since we are going to change q_ptr in the stream head.  Note that
7885 	 * insertq() doesn't wait for any syncq counts to drop to zero.
7886 	 */
7887 	sqlist.sqlist_head = NULL;
7888 	sqlist.sqlist_index = 0;
7889 	sqlist.sqlist_size = sizeof (sqlist_t);
7890 	sqlist_insert(&sqlist, _RD(stpdown->sd_wrq)->q_syncq);
7891 	strlock(stpdown, &sqlist);
7892 	strunlock(stpdown, &sqlist);
7893 
7894 	releaseq(passq);
7895 	return (passq);
7896 }
7897 
7898 /*
7899  * Let messages flow up into the mux by removing
7900  * the passq.
7901  */
7902 static void
link_rempassthru(queue_t * passq)7903 link_rempassthru(queue_t *passq)
7904 {
7905 	claimq(passq);
7906 	removeq(passq);
7907 	releaseq(passq);
7908 	freeq(passq);
7909 }
7910 
7911 /*
7912  * Wait for the condition variable pointed to by `cvp' to be signaled,
7913  * or for `tim' milliseconds to elapse, whichever comes first.  If `tim'
7914  * is negative, then there is no time limit.  If `nosigs' is non-zero,
7915  * then the wait will be non-interruptible.
7916  *
7917  * Returns >0 if signaled, 0 if interrupted, or -1 upon timeout.
7918  */
7919 clock_t
str_cv_wait(kcondvar_t * cvp,kmutex_t * mp,clock_t tim,int nosigs)7920 str_cv_wait(kcondvar_t *cvp, kmutex_t *mp, clock_t tim, int nosigs)
7921 {
7922 	clock_t ret;
7923 
7924 	if (tim < 0) {
7925 		if (nosigs) {
7926 			cv_wait(cvp, mp);
7927 			ret = 1;
7928 		} else {
7929 			ret = cv_wait_sig(cvp, mp);
7930 		}
7931 	} else if (tim > 0) {
7932 		/*
7933 		 * convert milliseconds to clock ticks
7934 		 */
7935 		if (nosigs) {
7936 			ret = cv_reltimedwait(cvp, mp,
7937 			    MSEC_TO_TICK_ROUNDUP(tim), TR_CLOCK_TICK);
7938 		} else {
7939 			ret = cv_reltimedwait_sig(cvp, mp,
7940 			    MSEC_TO_TICK_ROUNDUP(tim), TR_CLOCK_TICK);
7941 		}
7942 	} else {
7943 		ret = -1;
7944 	}
7945 	return (ret);
7946 }
7947 
7948 /*
7949  * Wait until the stream head can determine if it is at the mark but
7950  * don't wait forever to prevent a race condition between the "mark" state
7951  * in the stream head and any mark state in the caller/user of this routine.
7952  *
7953  * This is used by sockets and for a socket it would be incorrect
7954  * to return a failure for SIOCATMARK when there is no data in the receive
7955  * queue and the marked urgent data is traveling up the stream.
7956  *
7957  * This routine waits until the mark is known by waiting for one of these
7958  * three events:
7959  *	The stream head read queue becoming non-empty (including an EOF).
7960  *	The STRATMARK flag being set (due to a MSGMARKNEXT message).
7961  *	The STRNOTATMARK flag being set (which indicates that the transport
7962  *	has sent a MSGNOTMARKNEXT message to indicate that it is not at
7963  *	the mark).
7964  *
7965  * The routine returns 1 if the stream is at the mark; 0 if it can
7966  * be determined that the stream is not at the mark.
7967  * If the wait times out and it can't determine
7968  * whether or not the stream might be at the mark the routine will return -1.
7969  *
7970  * Note: This routine should only be used when a mark is pending i.e.,
7971  * in the socket case the SIGURG has been posted.
7972  * Note2: This can not wakeup just because synchronous streams indicate
7973  * that data is available since it is not possible to use the synchronous
7974  * streams interfaces to determine the b_flag value for the data queued below
7975  * the stream head.
7976  */
7977 int
strwaitmark(vnode_t * vp)7978 strwaitmark(vnode_t *vp)
7979 {
7980 	struct stdata *stp = vp->v_stream;
7981 	queue_t *rq = _RD(stp->sd_wrq);
7982 	int mark;
7983 
7984 	mutex_enter(&stp->sd_lock);
7985 	while (rq->q_first == NULL &&
7986 	    !(stp->sd_flag & (STRATMARK|STRNOTATMARK|STREOF))) {
7987 		stp->sd_flag |= RSLEEP;
7988 
7989 		/* Wait for 100 milliseconds for any state change. */
7990 		if (str_cv_wait(&rq->q_wait, &stp->sd_lock, 100, 1) == -1) {
7991 			mutex_exit(&stp->sd_lock);
7992 			return (-1);
7993 		}
7994 	}
7995 	if (stp->sd_flag & STRATMARK)
7996 		mark = 1;
7997 	else if (rq->q_first != NULL && (rq->q_first->b_flag & MSGMARK))
7998 		mark = 1;
7999 	else
8000 		mark = 0;
8001 
8002 	mutex_exit(&stp->sd_lock);
8003 	return (mark);
8004 }
8005 
8006 /*
8007  * Set a read side error. If persist is set change the socket error
8008  * to persistent. If errfunc is set install the function as the exported
8009  * error handler.
8010  */
8011 void
strsetrerror(vnode_t * vp,int error,int persist,errfunc_t errfunc)8012 strsetrerror(vnode_t *vp, int error, int persist, errfunc_t errfunc)
8013 {
8014 	struct stdata *stp = vp->v_stream;
8015 
8016 	mutex_enter(&stp->sd_lock);
8017 	stp->sd_rerror = error;
8018 	if (error == 0 && errfunc == NULL)
8019 		stp->sd_flag &= ~STRDERR;
8020 	else
8021 		stp->sd_flag |= STRDERR;
8022 	if (persist) {
8023 		stp->sd_flag &= ~STRDERRNONPERSIST;
8024 	} else {
8025 		stp->sd_flag |= STRDERRNONPERSIST;
8026 	}
8027 	stp->sd_rderrfunc = errfunc;
8028 	if (error != 0 || errfunc != NULL) {
8029 		cv_broadcast(&_RD(stp->sd_wrq)->q_wait);	/* readers */
8030 		cv_broadcast(&stp->sd_wrq->q_wait);		/* writers */
8031 		cv_broadcast(&stp->sd_monitor);			/* ioctllers */
8032 
8033 		mutex_exit(&stp->sd_lock);
8034 		pollwakeup(&stp->sd_pollist, POLLERR);
8035 		mutex_enter(&stp->sd_lock);
8036 
8037 		if (stp->sd_sigflags & S_ERROR)
8038 			strsendsig(stp->sd_siglist, S_ERROR, 0, error);
8039 	}
8040 	mutex_exit(&stp->sd_lock);
8041 }
8042 
8043 /*
8044  * Set a write side error. If persist is set change the socket error
8045  * to persistent.
8046  */
8047 void
strsetwerror(vnode_t * vp,int error,int persist,errfunc_t errfunc)8048 strsetwerror(vnode_t *vp, int error, int persist, errfunc_t errfunc)
8049 {
8050 	struct stdata *stp = vp->v_stream;
8051 
8052 	mutex_enter(&stp->sd_lock);
8053 	stp->sd_werror = error;
8054 	if (error == 0 && errfunc == NULL)
8055 		stp->sd_flag &= ~STWRERR;
8056 	else
8057 		stp->sd_flag |= STWRERR;
8058 	if (persist) {
8059 		stp->sd_flag &= ~STWRERRNONPERSIST;
8060 	} else {
8061 		stp->sd_flag |= STWRERRNONPERSIST;
8062 	}
8063 	stp->sd_wrerrfunc = errfunc;
8064 	if (error != 0 || errfunc != NULL) {
8065 		cv_broadcast(&_RD(stp->sd_wrq)->q_wait);	/* readers */
8066 		cv_broadcast(&stp->sd_wrq->q_wait);		/* writers */
8067 		cv_broadcast(&stp->sd_monitor);			/* ioctllers */
8068 
8069 		mutex_exit(&stp->sd_lock);
8070 		pollwakeup(&stp->sd_pollist, POLLERR);
8071 		mutex_enter(&stp->sd_lock);
8072 
8073 		if (stp->sd_sigflags & S_ERROR)
8074 			strsendsig(stp->sd_siglist, S_ERROR, 0, error);
8075 	}
8076 	mutex_exit(&stp->sd_lock);
8077 }
8078 
8079 /*
8080  * Make the stream return 0 (EOF) when all data has been read.
8081  * No effect on write side.
8082  */
8083 void
strseteof(vnode_t * vp,int eof)8084 strseteof(vnode_t *vp, int eof)
8085 {
8086 	struct stdata *stp = vp->v_stream;
8087 
8088 	mutex_enter(&stp->sd_lock);
8089 	if (!eof) {
8090 		stp->sd_flag &= ~STREOF;
8091 		mutex_exit(&stp->sd_lock);
8092 		return;
8093 	}
8094 	stp->sd_flag |= STREOF;
8095 	if (stp->sd_flag & RSLEEP) {
8096 		stp->sd_flag &= ~RSLEEP;
8097 		cv_broadcast(&_RD(stp->sd_wrq)->q_wait);
8098 	}
8099 
8100 	mutex_exit(&stp->sd_lock);
8101 	pollwakeup(&stp->sd_pollist, POLLIN|POLLRDNORM);
8102 	mutex_enter(&stp->sd_lock);
8103 
8104 	if (stp->sd_sigflags & (S_INPUT|S_RDNORM))
8105 		strsendsig(stp->sd_siglist, S_INPUT|S_RDNORM, 0, 0);
8106 	mutex_exit(&stp->sd_lock);
8107 }
8108 
8109 void
strflushrq(vnode_t * vp,int flag)8110 strflushrq(vnode_t *vp, int flag)
8111 {
8112 	struct stdata *stp = vp->v_stream;
8113 
8114 	mutex_enter(&stp->sd_lock);
8115 	flushq(_RD(stp->sd_wrq), flag);
8116 	mutex_exit(&stp->sd_lock);
8117 }
8118 
8119 void
strsetrputhooks(vnode_t * vp,uint_t flags,msgfunc_t protofunc,msgfunc_t miscfunc)8120 strsetrputhooks(vnode_t *vp, uint_t flags,
8121     msgfunc_t protofunc, msgfunc_t miscfunc)
8122 {
8123 	struct stdata *stp = vp->v_stream;
8124 
8125 	mutex_enter(&stp->sd_lock);
8126 
8127 	if (protofunc == NULL)
8128 		stp->sd_rprotofunc = strrput_proto;
8129 	else
8130 		stp->sd_rprotofunc = protofunc;
8131 
8132 	if (miscfunc == NULL)
8133 		stp->sd_rmiscfunc = strrput_misc;
8134 	else
8135 		stp->sd_rmiscfunc = miscfunc;
8136 
8137 	if (flags & SH_CONSOL_DATA)
8138 		stp->sd_rput_opt |= SR_CONSOL_DATA;
8139 	else
8140 		stp->sd_rput_opt &= ~SR_CONSOL_DATA;
8141 
8142 	if (flags & SH_SIGALLDATA)
8143 		stp->sd_rput_opt |= SR_SIGALLDATA;
8144 	else
8145 		stp->sd_rput_opt &= ~SR_SIGALLDATA;
8146 
8147 	if (flags & SH_IGN_ZEROLEN)
8148 		stp->sd_rput_opt |= SR_IGN_ZEROLEN;
8149 	else
8150 		stp->sd_rput_opt &= ~SR_IGN_ZEROLEN;
8151 
8152 	mutex_exit(&stp->sd_lock);
8153 }
8154 
8155 void
strsetwputhooks(vnode_t * vp,uint_t flags,clock_t closetime)8156 strsetwputhooks(vnode_t *vp, uint_t flags, clock_t closetime)
8157 {
8158 	struct stdata *stp = vp->v_stream;
8159 
8160 	mutex_enter(&stp->sd_lock);
8161 	stp->sd_closetime = closetime;
8162 
8163 	if (flags & SH_SIGPIPE)
8164 		stp->sd_wput_opt |= SW_SIGPIPE;
8165 	else
8166 		stp->sd_wput_opt &= ~SW_SIGPIPE;
8167 	if (flags & SH_RECHECK_ERR)
8168 		stp->sd_wput_opt |= SW_RECHECK_ERR;
8169 	else
8170 		stp->sd_wput_opt &= ~SW_RECHECK_ERR;
8171 
8172 	mutex_exit(&stp->sd_lock);
8173 }
8174 
8175 void
strsetrwputdatahooks(vnode_t * vp,msgfunc_t rdatafunc,msgfunc_t wdatafunc)8176 strsetrwputdatahooks(vnode_t *vp, msgfunc_t rdatafunc, msgfunc_t wdatafunc)
8177 {
8178 	struct stdata *stp = vp->v_stream;
8179 
8180 	mutex_enter(&stp->sd_lock);
8181 
8182 	stp->sd_rputdatafunc = rdatafunc;
8183 	stp->sd_wputdatafunc = wdatafunc;
8184 
8185 	mutex_exit(&stp->sd_lock);
8186 }
8187 
8188 /* Used within framework when the queue is already locked */
8189 void
qenable_locked(queue_t * q)8190 qenable_locked(queue_t *q)
8191 {
8192 	stdata_t *stp = STREAM(q);
8193 
8194 	ASSERT(MUTEX_HELD(QLOCK(q)));
8195 
8196 	if (!q->q_qinfo->qi_srvp)
8197 		return;
8198 
8199 	/*
8200 	 * Do not place on run queue if already enabled or closing.
8201 	 */
8202 	if (q->q_flag & (QWCLOSE|QENAB))
8203 		return;
8204 
8205 	/*
8206 	 * mark queue enabled and place on run list if it is not already being
8207 	 * serviced. If it is serviced, the runservice() function will detect
8208 	 * that QENAB is set and call service procedure before clearing
8209 	 * QINSERVICE flag.
8210 	 */
8211 	q->q_flag |= QENAB;
8212 	if (q->q_flag & QINSERVICE)
8213 		return;
8214 
8215 	/* Record the time of qenable */
8216 	q->q_qtstamp = ddi_get_lbolt();
8217 
8218 	/*
8219 	 * Put the queue in the stp list and schedule it for background
8220 	 * processing if it is not already scheduled or if stream head does not
8221 	 * intent to process it in the foreground later by setting
8222 	 * STRS_WILLSERVICE flag.
8223 	 */
8224 	mutex_enter(&stp->sd_qlock);
8225 	/*
8226 	 * If there are already something on the list, stp flags should show
8227 	 * intention to drain it.
8228 	 */
8229 	IMPLY(STREAM_NEEDSERVICE(stp),
8230 	    (stp->sd_svcflags & (STRS_WILLSERVICE | STRS_SCHEDULED)));
8231 
8232 	ENQUEUE(q, stp->sd_qhead, stp->sd_qtail, q_link);
8233 	stp->sd_nqueues++;
8234 
8235 	/*
8236 	 * If no one will drain this stream we are the first producer and
8237 	 * need to schedule it for background thread.
8238 	 */
8239 	if (!(stp->sd_svcflags & (STRS_WILLSERVICE | STRS_SCHEDULED))) {
8240 		/*
8241 		 * No one will service this stream later, so we have to
8242 		 * schedule it now.
8243 		 */
8244 		STRSTAT(stenables);
8245 		stp->sd_svcflags |= STRS_SCHEDULED;
8246 		stp->sd_servid = (void *)taskq_dispatch(streams_taskq,
8247 		    (task_func_t *)stream_service, stp, TQ_NOSLEEP|TQ_NOQUEUE);
8248 
8249 		if (stp->sd_servid == NULL) {
8250 			/*
8251 			 * Task queue failed so fail over to the backup
8252 			 * servicing thread.
8253 			 */
8254 			STRSTAT(taskqfails);
8255 			/*
8256 			 * It is safe to clear STRS_SCHEDULED flag because it
8257 			 * was set by this thread above.
8258 			 */
8259 			stp->sd_svcflags &= ~STRS_SCHEDULED;
8260 
8261 			/*
8262 			 * Failover scheduling is protected by service_queue
8263 			 * lock.
8264 			 */
8265 			mutex_enter(&service_queue);
8266 			ASSERT((stp->sd_qhead == q) && (stp->sd_qtail == q));
8267 			ASSERT(q->q_link == NULL);
8268 			/*
8269 			 * Append the queue to qhead/qtail list.
8270 			 */
8271 			if (qhead == NULL)
8272 				qhead = q;
8273 			else
8274 				qtail->q_link = q;
8275 			qtail = q;
8276 			/*
8277 			 * Clear stp queue list.
8278 			 */
8279 			stp->sd_qhead = stp->sd_qtail = NULL;
8280 			stp->sd_nqueues = 0;
8281 			/*
8282 			 * Wakeup background queue processing thread.
8283 			 */
8284 			cv_signal(&services_to_run);
8285 			mutex_exit(&service_queue);
8286 		}
8287 	}
8288 	mutex_exit(&stp->sd_qlock);
8289 }
8290 
8291 static void
queue_service(queue_t * q)8292 queue_service(queue_t *q)
8293 {
8294 	/*
8295 	 * The queue in the list should have
8296 	 * QENAB flag set and should not have
8297 	 * QINSERVICE flag set. QINSERVICE is
8298 	 * set when the queue is dequeued and
8299 	 * qenable_locked doesn't enqueue a
8300 	 * queue with QINSERVICE set.
8301 	 */
8302 
8303 	ASSERT(!(q->q_flag & QINSERVICE));
8304 	ASSERT((q->q_flag & QENAB));
8305 	mutex_enter(QLOCK(q));
8306 	q->q_flag &= ~QENAB;
8307 	q->q_flag |= QINSERVICE;
8308 	mutex_exit(QLOCK(q));
8309 	runservice(q);
8310 }
8311 
8312 static void
syncq_service(syncq_t * sq)8313 syncq_service(syncq_t *sq)
8314 {
8315 	STRSTAT(syncqservice);
8316 	mutex_enter(SQLOCK(sq));
8317 	ASSERT(!(sq->sq_svcflags & SQ_SERVICE));
8318 	ASSERT(sq->sq_servcount != 0);
8319 	ASSERT(sq->sq_next == NULL);
8320 
8321 	/* if we came here from the background thread, clear the flag */
8322 	if (sq->sq_svcflags & SQ_BGTHREAD)
8323 		sq->sq_svcflags &= ~SQ_BGTHREAD;
8324 
8325 	/* let drain_syncq know that it's being called in the background */
8326 	sq->sq_svcflags |= SQ_SERVICE;
8327 	drain_syncq(sq);
8328 }
8329 
8330 static void
qwriter_outer_service(syncq_t * outer)8331 qwriter_outer_service(syncq_t *outer)
8332 {
8333 	/*
8334 	 * Note that SQ_WRITER is used on the outer perimeter
8335 	 * to signal that a qwriter(OUTER) is either investigating
8336 	 * running or that it is actually running a function.
8337 	 */
8338 	outer_enter(outer, SQ_BLOCKED|SQ_WRITER);
8339 
8340 	/*
8341 	 * All inner syncq are empty and have SQ_WRITER set
8342 	 * to block entering the outer perimeter.
8343 	 *
8344 	 * We do not need to explicitly call write_now since
8345 	 * outer_exit does it for us.
8346 	 */
8347 	outer_exit(outer);
8348 }
8349 
8350 static void
mblk_free(mblk_t * mp)8351 mblk_free(mblk_t *mp)
8352 {
8353 	dblk_t *dbp = mp->b_datap;
8354 	frtn_t *frp = dbp->db_frtnp;
8355 
8356 	mp->b_next = NULL;
8357 	if (dbp->db_fthdr != NULL)
8358 		str_ftfree(dbp);
8359 
8360 	ASSERT(dbp->db_fthdr == NULL);
8361 	frp->free_func(frp->free_arg);
8362 	ASSERT(dbp->db_mblk == mp);
8363 
8364 	if (dbp->db_credp != NULL) {
8365 		crfree(dbp->db_credp);
8366 		dbp->db_credp = NULL;
8367 	}
8368 	dbp->db_cpid = -1;
8369 	dbp->db_struioflag = 0;
8370 	dbp->db_struioun.cksum.flags = 0;
8371 
8372 	kmem_cache_free(dbp->db_cache, dbp);
8373 }
8374 
8375 /*
8376  * Background processing of the stream queue list.
8377  */
8378 static void
stream_service(stdata_t * stp)8379 stream_service(stdata_t *stp)
8380 {
8381 	queue_t *q;
8382 
8383 	mutex_enter(&stp->sd_qlock);
8384 
8385 	STR_SERVICE(stp, q);
8386 
8387 	stp->sd_svcflags &= ~STRS_SCHEDULED;
8388 	stp->sd_servid = NULL;
8389 	cv_signal(&stp->sd_qcv);
8390 	mutex_exit(&stp->sd_qlock);
8391 }
8392 
8393 /*
8394  * Foreground processing of the stream queue list.
8395  */
8396 void
stream_runservice(stdata_t * stp)8397 stream_runservice(stdata_t *stp)
8398 {
8399 	queue_t *q;
8400 
8401 	mutex_enter(&stp->sd_qlock);
8402 	STRSTAT(rservice);
8403 	/*
8404 	 * We are going to drain this stream queue list, so qenable_locked will
8405 	 * not schedule it until we finish.
8406 	 */
8407 	stp->sd_svcflags |= STRS_WILLSERVICE;
8408 
8409 	STR_SERVICE(stp, q);
8410 
8411 	stp->sd_svcflags &= ~STRS_WILLSERVICE;
8412 	mutex_exit(&stp->sd_qlock);
8413 }
8414 
8415 void
stream_willservice(stdata_t * stp)8416 stream_willservice(stdata_t *stp)
8417 {
8418 	mutex_enter(&stp->sd_qlock);
8419 	stp->sd_svcflags |= STRS_WILLSERVICE;
8420 	mutex_exit(&stp->sd_qlock);
8421 }
8422 
8423 /*
8424  * Replace the cred currently in the mblk with a different one.
8425  * Also update db_cpid.
8426  */
8427 void
mblk_setcred(mblk_t * mp,cred_t * cr,pid_t cpid)8428 mblk_setcred(mblk_t *mp, cred_t *cr, pid_t cpid)
8429 {
8430 	dblk_t *dbp = mp->b_datap;
8431 	cred_t *ocr = dbp->db_credp;
8432 
8433 	ASSERT(cr != NULL);
8434 
8435 	if (cr != ocr) {
8436 		crhold(dbp->db_credp = cr);
8437 		if (ocr != NULL)
8438 			crfree(ocr);
8439 	}
8440 	/* Don't overwrite with NOPID */
8441 	if (cpid != NOPID)
8442 		dbp->db_cpid = cpid;
8443 }
8444 
8445 /*
8446  * If the src message has a cred, then replace the cred currently in the mblk
8447  * with it.
8448  * Also update db_cpid.
8449  */
8450 void
mblk_copycred(mblk_t * mp,const mblk_t * src)8451 mblk_copycred(mblk_t *mp, const mblk_t *src)
8452 {
8453 	dblk_t *dbp = mp->b_datap;
8454 	cred_t *cr, *ocr;
8455 	pid_t cpid;
8456 
8457 	cr = msg_getcred(src, &cpid);
8458 	if (cr == NULL)
8459 		return;
8460 
8461 	ocr = dbp->db_credp;
8462 	if (cr != ocr) {
8463 		crhold(dbp->db_credp = cr);
8464 		if (ocr != NULL)
8465 			crfree(ocr);
8466 	}
8467 	/* Don't overwrite with NOPID */
8468 	if (cpid != NOPID)
8469 		dbp->db_cpid = cpid;
8470 }
8471 
8472 void
lso_info_set(mblk_t * mp,uint32_t mss,uint32_t flags)8473 lso_info_set(mblk_t *mp, uint32_t mss, uint32_t flags)
8474 {
8475 	ASSERT(DB_TYPE(mp) == M_DATA);
8476 	ASSERT((flags & ~HW_LSO_FLAGS) == 0);
8477 
8478 	/* Set the flags */
8479 	DB_LSOFLAGS(mp) |= flags;
8480 	DB_LSOMSS(mp) = mss;
8481 }
8482 
8483 void
lso_info_cleanup(mblk_t * mp)8484 lso_info_cleanup(mblk_t *mp)
8485 {
8486 	ASSERT(DB_TYPE(mp) == M_DATA);
8487 
8488 	/* Clear the flags */
8489 	DB_LSOFLAGS(mp) &= ~HW_LSO_FLAGS;
8490 	DB_LSOMSS(mp) = 0;
8491 }
8492 
8493 /*
8494  * Checksum buffer *bp for len bytes with psum partial checksum,
8495  * or 0 if none, and return the 16 bit partial checksum.
8496  */
8497 unsigned
bcksum(uchar_t * bp,int len,unsigned int psum)8498 bcksum(uchar_t *bp, int len, unsigned int psum)
8499 {
8500 	int odd = len & 1;
8501 	extern unsigned int ip_ocsum();
8502 
8503 	if (((intptr_t)bp & 1) == 0 && !odd) {
8504 		/*
8505 		 * Bp is 16 bit aligned and len is multiple of 16 bit word.
8506 		 */
8507 		return (ip_ocsum((ushort_t *)bp, len >> 1, psum));
8508 	}
8509 	if (((intptr_t)bp & 1) != 0) {
8510 		/*
8511 		 * Bp isn't 16 bit aligned.
8512 		 */
8513 		unsigned int tsum;
8514 
8515 #ifdef _LITTLE_ENDIAN
8516 		psum += *bp;
8517 #else
8518 		psum += *bp << 8;
8519 #endif
8520 		len--;
8521 		bp++;
8522 		tsum = ip_ocsum((ushort_t *)bp, len >> 1, 0);
8523 		psum += (tsum << 8) & 0xffff | (tsum >> 8);
8524 		if (len & 1) {
8525 			bp += len - 1;
8526 #ifdef _LITTLE_ENDIAN
8527 			psum += *bp << 8;
8528 #else
8529 			psum += *bp;
8530 #endif
8531 		}
8532 	} else {
8533 		/*
8534 		 * Bp is 16 bit aligned.
8535 		 */
8536 		psum = ip_ocsum((ushort_t *)bp, len >> 1, psum);
8537 		if (odd) {
8538 			bp += len - 1;
8539 #ifdef _LITTLE_ENDIAN
8540 			psum += *bp;
8541 #else
8542 			psum += *bp << 8;
8543 #endif
8544 		}
8545 	}
8546 	/*
8547 	 * Normalize psum to 16 bits before returning the new partial
8548 	 * checksum. The max psum value before normalization is 0x3FDFE.
8549 	 */
8550 	return ((psum >> 16) + (psum & 0xFFFF));
8551 }
8552 
8553 void
freemsgchain(mblk_t * mp)8554 freemsgchain(mblk_t *mp)
8555 {
8556 	mblk_t	*next;
8557 
8558 	while (mp != NULL) {
8559 		next = mp->b_next;
8560 		mp->b_next = NULL;
8561 
8562 		freemsg(mp);
8563 		mp = next;
8564 	}
8565 }
8566 
8567 mblk_t *
copymsgchain(mblk_t * mp)8568 copymsgchain(mblk_t *mp)
8569 {
8570 	mblk_t	*nmp = NULL;
8571 	mblk_t	**nmpp = &nmp;
8572 
8573 	for (; mp != NULL; mp = mp->b_next) {
8574 		if ((*nmpp = copymsg(mp)) == NULL) {
8575 			freemsgchain(nmp);
8576 			return (NULL);
8577 		}
8578 
8579 		nmpp = &((*nmpp)->b_next);
8580 	}
8581 
8582 	return (nmp);
8583 }
8584 
8585 /* NOTE: Do not add code after this point. */
8586 #undef QLOCK
8587 
8588 /*
8589  * Replacement for QLOCK macro for those that can't use it.
8590  */
8591 kmutex_t *
QLOCK(queue_t * q)8592 QLOCK(queue_t *q)
8593 {
8594 	return (&(q)->q_lock);
8595 }
8596 
8597 /*
8598  * Dummy runqueues/queuerun functions functions for backwards compatibility.
8599  */
8600 #undef runqueues
8601 void
runqueues(void)8602 runqueues(void)
8603 {
8604 }
8605 
8606 #undef queuerun
8607 void
queuerun(void)8608 queuerun(void)
8609 {
8610 }
8611 
8612 /*
8613  * Initialize the STR stack instance, which tracks autopush and persistent
8614  * links.
8615  */
8616 /* ARGSUSED */
8617 static void *
str_stack_init(netstackid_t stackid,netstack_t * ns)8618 str_stack_init(netstackid_t stackid, netstack_t *ns)
8619 {
8620 	str_stack_t	*ss;
8621 	int i;
8622 
8623 	ss = (str_stack_t *)kmem_zalloc(sizeof (*ss), KM_SLEEP);
8624 	ss->ss_netstack = ns;
8625 
8626 	/*
8627 	 * set up autopush
8628 	 */
8629 	sad_initspace(ss);
8630 
8631 	/*
8632 	 * set up mux_node structures.
8633 	 */
8634 	ss->ss_devcnt = devcnt;	/* In case it should change before free */
8635 	ss->ss_mux_nodes = kmem_zalloc((sizeof (struct mux_node) *
8636 	    ss->ss_devcnt), KM_SLEEP);
8637 	for (i = 0; i < ss->ss_devcnt; i++)
8638 		ss->ss_mux_nodes[i].mn_imaj = i;
8639 	return (ss);
8640 }
8641 
8642 /*
8643  * Note: run at zone shutdown and not destroy so that the PLINKs are
8644  * gone by the time other cleanup happens from the destroy callbacks.
8645  */
8646 static void
str_stack_shutdown(netstackid_t stackid,void * arg)8647 str_stack_shutdown(netstackid_t stackid, void *arg)
8648 {
8649 	str_stack_t *ss = (str_stack_t *)arg;
8650 	int i;
8651 	cred_t *cr;
8652 
8653 	cr = zone_get_kcred(netstackid_to_zoneid(stackid));
8654 	ASSERT(cr != NULL);
8655 
8656 	/* Undo all the I_PLINKs for this zone */
8657 	for (i = 0; i < ss->ss_devcnt; i++) {
8658 		struct mux_edge		*ep;
8659 		ldi_handle_t		lh;
8660 		ldi_ident_t		li;
8661 		int			ret;
8662 		int			rval;
8663 		dev_t			rdev;
8664 
8665 		ep = ss->ss_mux_nodes[i].mn_outp;
8666 		if (ep == NULL)
8667 			continue;
8668 		ret = ldi_ident_from_major((major_t)i, &li);
8669 		if (ret != 0) {
8670 			continue;
8671 		}
8672 		rdev = ep->me_dev;
8673 		ret = ldi_open_by_dev(&rdev, OTYP_CHR, FREAD|FWRITE,
8674 		    cr, &lh, li);
8675 		if (ret != 0) {
8676 			ldi_ident_release(li);
8677 			continue;
8678 		}
8679 
8680 		ret = ldi_ioctl(lh, I_PUNLINK, (intptr_t)MUXID_ALL, FKIOCTL,
8681 		    cr, &rval);
8682 		if (ret) {
8683 			(void) ldi_close(lh, FREAD|FWRITE, cr);
8684 			ldi_ident_release(li);
8685 			continue;
8686 		}
8687 		(void) ldi_close(lh, FREAD|FWRITE, cr);
8688 
8689 		/* Close layered handles */
8690 		ldi_ident_release(li);
8691 	}
8692 	crfree(cr);
8693 
8694 	sad_freespace(ss);
8695 
8696 	kmem_free(ss->ss_mux_nodes, sizeof (struct mux_node) * ss->ss_devcnt);
8697 	ss->ss_mux_nodes = NULL;
8698 }
8699 
8700 /*
8701  * Free the structure; str_stack_shutdown did the other cleanup work.
8702  */
8703 /* ARGSUSED */
8704 static void
str_stack_fini(netstackid_t stackid,void * arg)8705 str_stack_fini(netstackid_t stackid, void *arg)
8706 {
8707 	str_stack_t	*ss = (str_stack_t *)arg;
8708 
8709 	kmem_free(ss, sizeof (*ss));
8710 }
8711