1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /* Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T */
22 /* All Rights Reserved */
23
24
25 /*
26 * Copyright 2010 Sun Microsystems, Inc. All rights reserved.
27 * Use is subject to license terms.
28 */
29
30 #include <sys/types.h>
31 #include <sys/sysmacros.h>
32 #include <sys/param.h>
33 #include <sys/errno.h>
34 #include <sys/signal.h>
35 #include <sys/proc.h>
36 #include <sys/conf.h>
37 #include <sys/cred.h>
38 #include <sys/user.h>
39 #include <sys/vnode.h>
40 #include <sys/file.h>
41 #include <sys/session.h>
42 #include <sys/stream.h>
43 #include <sys/strsubr.h>
44 #include <sys/stropts.h>
45 #include <sys/poll.h>
46 #include <sys/systm.h>
47 #include <sys/cpuvar.h>
48 #include <sys/uio.h>
49 #include <sys/cmn_err.h>
50 #include <sys/priocntl.h>
51 #include <sys/procset.h>
52 #include <sys/vmem.h>
53 #include <sys/bitmap.h>
54 #include <sys/kmem.h>
55 #include <sys/siginfo.h>
56 #include <sys/vtrace.h>
57 #include <sys/callb.h>
58 #include <sys/debug.h>
59 #include <sys/modctl.h>
60 #include <sys/vmsystm.h>
61 #include <vm/page.h>
62 #include <sys/atomic.h>
63 #include <sys/suntpi.h>
64 #include <sys/strlog.h>
65 #include <sys/promif.h>
66 #include <sys/project.h>
67 #include <sys/vm.h>
68 #include <sys/taskq.h>
69 #include <sys/sunddi.h>
70 #include <sys/sunldi_impl.h>
71 #include <sys/strsun.h>
72 #include <sys/isa_defs.h>
73 #include <sys/multidata.h>
74 #include <sys/pattr.h>
75 #include <sys/strft.h>
76 #include <sys/fs/snode.h>
77 #include <sys/zone.h>
78 #include <sys/open.h>
79 #include <sys/sunldi.h>
80 #include <sys/sad.h>
81 #include <sys/netstack.h>
82
83 #define O_SAMESTR(q) (((q)->q_next) && \
84 (((q)->q_flag & QREADR) == ((q)->q_next->q_flag & QREADR)))
85
86 /*
87 * WARNING:
88 * The variables and routines in this file are private, belonging
89 * to the STREAMS subsystem. These should not be used by modules
90 * or drivers. Compatibility will not be guaranteed.
91 */
92
93 /*
94 * Id value used to distinguish between different multiplexor links.
95 */
96 static int32_t lnk_id = 0;
97
98 #define STREAMS_LOPRI MINCLSYSPRI
99 static pri_t streams_lopri = STREAMS_LOPRI;
100
101 #define STRSTAT(x) (str_statistics.x.value.ui64++)
102 typedef struct str_stat {
103 kstat_named_t sqenables;
104 kstat_named_t stenables;
105 kstat_named_t syncqservice;
106 kstat_named_t freebs;
107 kstat_named_t qwr_outer;
108 kstat_named_t rservice;
109 kstat_named_t strwaits;
110 kstat_named_t taskqfails;
111 kstat_named_t bufcalls;
112 kstat_named_t qhelps;
113 kstat_named_t qremoved;
114 kstat_named_t sqremoved;
115 kstat_named_t bcwaits;
116 kstat_named_t sqtoomany;
117 } str_stat_t;
118
119 static str_stat_t str_statistics = {
120 { "sqenables", KSTAT_DATA_UINT64 },
121 { "stenables", KSTAT_DATA_UINT64 },
122 { "syncqservice", KSTAT_DATA_UINT64 },
123 { "freebs", KSTAT_DATA_UINT64 },
124 { "qwr_outer", KSTAT_DATA_UINT64 },
125 { "rservice", KSTAT_DATA_UINT64 },
126 { "strwaits", KSTAT_DATA_UINT64 },
127 { "taskqfails", KSTAT_DATA_UINT64 },
128 { "bufcalls", KSTAT_DATA_UINT64 },
129 { "qhelps", KSTAT_DATA_UINT64 },
130 { "qremoved", KSTAT_DATA_UINT64 },
131 { "sqremoved", KSTAT_DATA_UINT64 },
132 { "bcwaits", KSTAT_DATA_UINT64 },
133 { "sqtoomany", KSTAT_DATA_UINT64 },
134 };
135
136 static kstat_t *str_kstat;
137
138 /*
139 * qrunflag was used previously to control background scheduling of queues. It
140 * is not used anymore, but kept here in case some module still wants to access
141 * it via qready() and setqsched macros.
142 */
143 char qrunflag; /* Unused */
144
145 /*
146 * Most of the streams scheduling is done via task queues. Task queues may fail
147 * for non-sleep dispatches, so there are two backup threads servicing failed
148 * requests for queues and syncqs. Both of these threads also service failed
149 * dispatches freebs requests. Queues are put in the list specified by `qhead'
150 * and `qtail' pointers, syncqs use `sqhead' and `sqtail' pointers and freebs
151 * requests are put into `freebs_list' which has no tail pointer. All three
152 * lists are protected by a single `service_queue' lock and use
153 * `services_to_run' condition variable for signaling background threads. Use of
154 * a single lock should not be a problem because it is only used under heavy
155 * loads when task queues start to fail and at that time it may be a good idea
156 * to throttle scheduling requests.
157 *
158 * NOTE: queues and syncqs should be scheduled by two separate threads because
159 * queue servicing may be blocked waiting for a syncq which may be also
160 * scheduled for background execution. This may create a deadlock when only one
161 * thread is used for both.
162 */
163
164 static taskq_t *streams_taskq; /* Used for most STREAMS scheduling */
165
166 static kmutex_t service_queue; /* protects all of servicing vars */
167 static kcondvar_t services_to_run; /* wake up background service thread */
168 static kcondvar_t syncqs_to_run; /* wake up background service thread */
169
170 /*
171 * List of queues scheduled for background processing due to lack of resources
172 * in the task queues. Protected by service_queue lock;
173 */
174 static struct queue *qhead;
175 static struct queue *qtail;
176
177 /*
178 * Same list for syncqs
179 */
180 static syncq_t *sqhead;
181 static syncq_t *sqtail;
182
183 static mblk_t *freebs_list; /* list of buffers to free */
184
185 /*
186 * Backup threads for servicing queues and syncqs
187 */
188 kthread_t *streams_qbkgrnd_thread;
189 kthread_t *streams_sqbkgrnd_thread;
190
191 /*
192 * Bufcalls related variables.
193 */
194 struct bclist strbcalls; /* list of waiting bufcalls */
195 kmutex_t strbcall_lock; /* protects bufcall list (strbcalls) */
196 kcondvar_t strbcall_cv; /* Signaling when a bufcall is added */
197 kmutex_t bcall_monitor; /* sleep/wakeup style monitor */
198 kcondvar_t bcall_cv; /* wait 'till executing bufcall completes */
199 kthread_t *bc_bkgrnd_thread; /* Thread to service bufcall requests */
200
201 kmutex_t strresources; /* protects global resources */
202 kmutex_t muxifier; /* single-threads multiplexor creation */
203
204 static void *str_stack_init(netstackid_t stackid, netstack_t *ns);
205 static void str_stack_shutdown(netstackid_t stackid, void *arg);
206 static void str_stack_fini(netstackid_t stackid, void *arg);
207
208 /*
209 * run_queues is no longer used, but is kept in case some 3rd party
210 * module/driver decides to use it.
211 */
212 int run_queues = 0;
213
214 /*
215 * sq_max_size is the depth of the syncq (in number of messages) before
216 * qfill_syncq() starts QFULL'ing destination queues. As its primary
217 * consumer - IP is no longer D_MTPERMOD, but there may be other
218 * modules/drivers depend on this syncq flow control, we prefer to
219 * choose a large number as the default value. For potential
220 * performance gain, this value is tunable in /etc/system.
221 */
222 int sq_max_size = 10000;
223
224 /*
225 * The number of ciputctrl structures per syncq and stream we create when
226 * needed.
227 */
228 int n_ciputctrl;
229 int max_n_ciputctrl = 16;
230 /*
231 * If n_ciputctrl is < min_n_ciputctrl don't even create ciputctrl_cache.
232 */
233 int min_n_ciputctrl = 2;
234
235 /*
236 * Per-driver/module syncqs
237 * ========================
238 *
239 * For drivers/modules that use PERMOD or outer syncqs we keep a list of
240 * perdm structures, new entries being added (and new syncqs allocated) when
241 * setq() encounters a module/driver with a streamtab that it hasn't seen
242 * before.
243 * The reason for this mechanism is that some modules and drivers share a
244 * common streamtab and it is necessary for those modules and drivers to also
245 * share a common PERMOD syncq.
246 *
247 * perdm_list --> dm_str == streamtab_1
248 * dm_sq == syncq_1
249 * dm_ref
250 * dm_next --> dm_str == streamtab_2
251 * dm_sq == syncq_2
252 * dm_ref
253 * dm_next --> ... NULL
254 *
255 * The dm_ref field is incremented for each new driver/module that takes
256 * a reference to the perdm structure and hence shares the syncq.
257 * References are held in the fmodsw_impl_t structure for each STREAMS module
258 * or the dev_impl array (indexed by device major number) for each driver.
259 *
260 * perdm_list -> [dm_ref == 1] -> [dm_ref == 2] -> [dm_ref == 1] -> NULL
261 * ^ ^ ^ ^
262 * | ______________/ | |
263 * | / | |
264 * dev_impl: ...|x|y|... module A module B
265 *
266 * When a module/driver is unloaded the reference count is decremented and,
267 * when it falls to zero, the perdm structure is removed from the list and
268 * the syncq is freed (see rele_dm()).
269 */
270 perdm_t *perdm_list = NULL;
271 static krwlock_t perdm_rwlock;
272 cdevsw_impl_t *devimpl;
273
274 extern struct qinit strdata;
275 extern struct qinit stwdata;
276
277 static void runservice(queue_t *);
278 static void streams_bufcall_service(void);
279 static void streams_qbkgrnd_service(void);
280 static void streams_sqbkgrnd_service(void);
281 static syncq_t *new_syncq(void);
282 static void free_syncq(syncq_t *);
283 static void outer_insert(syncq_t *, syncq_t *);
284 static void outer_remove(syncq_t *, syncq_t *);
285 static void write_now(syncq_t *);
286 static void clr_qfull(queue_t *);
287 static void runbufcalls(void);
288 static void sqenable(syncq_t *);
289 static void sqfill_events(syncq_t *, queue_t *, mblk_t *, void (*)());
290 static void wait_q_syncq(queue_t *);
291 static void backenable_insertedq(queue_t *);
292
293 static void queue_service(queue_t *);
294 static void stream_service(stdata_t *);
295 static void syncq_service(syncq_t *);
296 static void qwriter_outer_service(syncq_t *);
297 static void mblk_free(mblk_t *);
298 #ifdef DEBUG
299 static int qprocsareon(queue_t *);
300 #endif
301
302 static void set_nfsrv_ptr(queue_t *, queue_t *, queue_t *, queue_t *);
303 static void reset_nfsrv_ptr(queue_t *, queue_t *);
304 void set_qfull(queue_t *);
305
306 static void sq_run_events(syncq_t *);
307 static int propagate_syncq(queue_t *);
308
309 static void blocksq(syncq_t *, ushort_t, int);
310 static void unblocksq(syncq_t *, ushort_t, int);
311 static int dropsq(syncq_t *, uint16_t);
312 static void emptysq(syncq_t *);
313 static sqlist_t *sqlist_alloc(struct stdata *, int);
314 static void sqlist_free(sqlist_t *);
315 static sqlist_t *sqlist_build(queue_t *, struct stdata *, boolean_t);
316 static void sqlist_insert(sqlist_t *, syncq_t *);
317 static void sqlist_insertall(sqlist_t *, queue_t *);
318
319 static void strsetuio(stdata_t *);
320
321 struct kmem_cache *stream_head_cache;
322 struct kmem_cache *queue_cache;
323 struct kmem_cache *syncq_cache;
324 struct kmem_cache *qband_cache;
325 struct kmem_cache *linkinfo_cache;
326 struct kmem_cache *ciputctrl_cache = NULL;
327
328 static linkinfo_t *linkinfo_list;
329
330 /* Global esballoc throttling queue */
331 static esb_queue_t system_esbq;
332
333 /* Array of esballoc throttling queues, of length esbq_nelem */
334 static esb_queue_t *volatile system_esbq_array;
335 static int esbq_nelem;
336 static kmutex_t esbq_lock;
337 static int esbq_log2_cpus_per_q = 0;
338
339 /* Scale the system_esbq length by setting number of CPUs per queue. */
340 uint_t esbq_cpus_per_q = 1;
341
342 /*
343 * esballoc tunable parameters.
344 */
345 int esbq_max_qlen = 0x16; /* throttled queue length */
346 clock_t esbq_timeout = 0x8; /* timeout to process esb queue */
347
348 /*
349 * Routines to handle esballoc queueing.
350 */
351 static void esballoc_process_queue(esb_queue_t *);
352 static void esballoc_enqueue_mblk(mblk_t *);
353 static void esballoc_timer(void *);
354 static void esballoc_set_timer(esb_queue_t *, clock_t);
355 static void esballoc_mblk_free(mblk_t *);
356
357 /*
358 * Qinit structure and Module_info structures
359 * for passthru read and write queues
360 */
361
362 static void pass_wput(queue_t *, mblk_t *);
363 static queue_t *link_addpassthru(stdata_t *);
364 static void link_rempassthru(queue_t *);
365
366 struct module_info passthru_info = {
367 0,
368 "passthru",
369 0,
370 INFPSZ,
371 STRHIGH,
372 STRLOW
373 };
374
375 struct qinit passthru_rinit = {
376 (int (*)())putnext,
377 NULL,
378 NULL,
379 NULL,
380 NULL,
381 &passthru_info,
382 NULL
383 };
384
385 struct qinit passthru_winit = {
386 (int (*)()) pass_wput,
387 NULL,
388 NULL,
389 NULL,
390 NULL,
391 &passthru_info,
392 NULL
393 };
394
395 /*
396 * Verify correctness of list head/tail pointers.
397 */
398 #define LISTCHECK(head, tail, link) { \
399 EQUIV(head, tail); \
400 IMPLY(tail != NULL, tail->link == NULL); \
401 }
402
403 /*
404 * Enqueue a list element `el' in the end of a list denoted by `head' and `tail'
405 * using a `link' field.
406 */
407 #define ENQUEUE(el, head, tail, link) { \
408 ASSERT(el->link == NULL); \
409 LISTCHECK(head, tail, link); \
410 if (head == NULL) \
411 head = el; \
412 else \
413 tail->link = el; \
414 tail = el; \
415 }
416
417 /*
418 * Dequeue the first element of the list denoted by `head' and `tail' pointers
419 * using a `link' field and put result into `el'.
420 */
421 #define DQ(el, head, tail, link) { \
422 LISTCHECK(head, tail, link); \
423 el = head; \
424 if (head != NULL) { \
425 head = head->link; \
426 if (head == NULL) \
427 tail = NULL; \
428 el->link = NULL; \
429 } \
430 }
431
432 /*
433 * Remove `el' from the list using `chase' and `curr' pointers and return result
434 * in `succeed'.
435 */
436 #define RMQ(el, head, tail, link, chase, curr, succeed) { \
437 LISTCHECK(head, tail, link); \
438 chase = NULL; \
439 succeed = 0; \
440 for (curr = head; (curr != el) && (curr != NULL); curr = curr->link) \
441 chase = curr; \
442 if (curr != NULL) { \
443 succeed = 1; \
444 ASSERT(curr == el); \
445 if (chase != NULL) \
446 chase->link = curr->link; \
447 else \
448 head = curr->link; \
449 curr->link = NULL; \
450 if (curr == tail) \
451 tail = chase; \
452 } \
453 LISTCHECK(head, tail, link); \
454 }
455
456 /* Handling of delayed messages on the inner syncq. */
457
458 /*
459 * DEBUG versions should use function versions (to simplify tracing) and
460 * non-DEBUG kernels should use macro versions.
461 */
462
463 /*
464 * Put a queue on the syncq list of queues.
465 * Assumes SQLOCK held.
466 */
467 #define SQPUT_Q(sq, qp) \
468 { \
469 ASSERT(MUTEX_HELD(SQLOCK(sq))); \
470 if (!(qp->q_sqflags & Q_SQQUEUED)) { \
471 /* The queue should not be linked anywhere */ \
472 ASSERT((qp->q_sqprev == NULL) && (qp->q_sqnext == NULL)); \
473 /* Head and tail may only be NULL simultaneously */ \
474 EQUIV(sq->sq_head, sq->sq_tail); \
475 /* Queue may be only enqueued on its syncq */ \
476 ASSERT(sq == qp->q_syncq); \
477 /* Check the correctness of SQ_MESSAGES flag */ \
478 EQUIV(sq->sq_head, (sq->sq_flags & SQ_MESSAGES)); \
479 /* Sanity check first/last elements of the list */ \
480 IMPLY(sq->sq_head != NULL, sq->sq_head->q_sqprev == NULL);\
481 IMPLY(sq->sq_tail != NULL, sq->sq_tail->q_sqnext == NULL);\
482 /* \
483 * Sanity check of priority field: empty queue should \
484 * have zero priority \
485 * and nqueues equal to zero. \
486 */ \
487 IMPLY(sq->sq_head == NULL, sq->sq_pri == 0); \
488 /* Sanity check of sq_nqueues field */ \
489 EQUIV(sq->sq_head, sq->sq_nqueues); \
490 if (sq->sq_head == NULL) { \
491 sq->sq_head = sq->sq_tail = qp; \
492 sq->sq_flags |= SQ_MESSAGES; \
493 } else if (qp->q_spri == 0) { \
494 qp->q_sqprev = sq->sq_tail; \
495 sq->sq_tail->q_sqnext = qp; \
496 sq->sq_tail = qp; \
497 } else { \
498 /* \
499 * Put this queue in priority order: higher \
500 * priority gets closer to the head. \
501 */ \
502 queue_t **qpp = &sq->sq_tail; \
503 queue_t *qnext = NULL; \
504 \
505 while (*qpp != NULL && qp->q_spri > (*qpp)->q_spri) { \
506 qnext = *qpp; \
507 qpp = &(*qpp)->q_sqprev; \
508 } \
509 qp->q_sqnext = qnext; \
510 qp->q_sqprev = *qpp; \
511 if (*qpp != NULL) { \
512 (*qpp)->q_sqnext = qp; \
513 } else { \
514 sq->sq_head = qp; \
515 sq->sq_pri = sq->sq_head->q_spri; \
516 } \
517 *qpp = qp; \
518 } \
519 qp->q_sqflags |= Q_SQQUEUED; \
520 qp->q_sqtstamp = ddi_get_lbolt(); \
521 sq->sq_nqueues++; \
522 } \
523 }
524
525 /*
526 * Remove a queue from the syncq list
527 * Assumes SQLOCK held.
528 */
529 #define SQRM_Q(sq, qp) \
530 { \
531 ASSERT(MUTEX_HELD(SQLOCK(sq))); \
532 ASSERT(qp->q_sqflags & Q_SQQUEUED); \
533 ASSERT(sq->sq_head != NULL && sq->sq_tail != NULL); \
534 ASSERT((sq->sq_flags & SQ_MESSAGES) != 0); \
535 /* Check that the queue is actually in the list */ \
536 ASSERT(qp->q_sqnext != NULL || sq->sq_tail == qp); \
537 ASSERT(qp->q_sqprev != NULL || sq->sq_head == qp); \
538 ASSERT(sq->sq_nqueues != 0); \
539 if (qp->q_sqprev == NULL) { \
540 /* First queue on list, make head q_sqnext */ \
541 sq->sq_head = qp->q_sqnext; \
542 } else { \
543 /* Make prev->next == next */ \
544 qp->q_sqprev->q_sqnext = qp->q_sqnext; \
545 } \
546 if (qp->q_sqnext == NULL) { \
547 /* Last queue on list, make tail sqprev */ \
548 sq->sq_tail = qp->q_sqprev; \
549 } else { \
550 /* Make next->prev == prev */ \
551 qp->q_sqnext->q_sqprev = qp->q_sqprev; \
552 } \
553 /* clear out references on this queue */ \
554 qp->q_sqprev = qp->q_sqnext = NULL; \
555 qp->q_sqflags &= ~Q_SQQUEUED; \
556 /* If there is nothing queued, clear SQ_MESSAGES */ \
557 if (sq->sq_head != NULL) { \
558 sq->sq_pri = sq->sq_head->q_spri; \
559 } else { \
560 sq->sq_flags &= ~SQ_MESSAGES; \
561 sq->sq_pri = 0; \
562 } \
563 sq->sq_nqueues--; \
564 ASSERT(sq->sq_head != NULL || sq->sq_evhead != NULL || \
565 (sq->sq_flags & SQ_QUEUED) == 0); \
566 }
567
568 /* Hide the definition from the header file. */
569 #ifdef SQPUT_MP
570 #undef SQPUT_MP
571 #endif
572
573 /*
574 * Put a message on the queue syncq.
575 * Assumes QLOCK held.
576 */
577 #define SQPUT_MP(qp, mp) \
578 { \
579 ASSERT(MUTEX_HELD(QLOCK(qp))); \
580 ASSERT(qp->q_sqhead == NULL || \
581 (qp->q_sqtail != NULL && \
582 qp->q_sqtail->b_next == NULL)); \
583 qp->q_syncqmsgs++; \
584 ASSERT(qp->q_syncqmsgs != 0); /* Wraparound */ \
585 if (qp->q_sqhead == NULL) { \
586 qp->q_sqhead = qp->q_sqtail = mp; \
587 } else { \
588 qp->q_sqtail->b_next = mp; \
589 qp->q_sqtail = mp; \
590 } \
591 ASSERT(qp->q_syncqmsgs > 0); \
592 set_qfull(qp); \
593 }
594
595 #define SQ_PUTCOUNT_SETFAST_LOCKED(sq) { \
596 ASSERT(MUTEX_HELD(SQLOCK(sq))); \
597 if ((sq)->sq_ciputctrl != NULL) { \
598 int i; \
599 int nlocks = (sq)->sq_nciputctrl; \
600 ciputctrl_t *cip = (sq)->sq_ciputctrl; \
601 ASSERT((sq)->sq_type & SQ_CIPUT); \
602 for (i = 0; i <= nlocks; i++) { \
603 ASSERT(MUTEX_HELD(&cip[i].ciputctrl_lock)); \
604 cip[i].ciputctrl_count |= SQ_FASTPUT; \
605 } \
606 } \
607 }
608
609
610 #define SQ_PUTCOUNT_CLRFAST_LOCKED(sq) { \
611 ASSERT(MUTEX_HELD(SQLOCK(sq))); \
612 if ((sq)->sq_ciputctrl != NULL) { \
613 int i; \
614 int nlocks = (sq)->sq_nciputctrl; \
615 ciputctrl_t *cip = (sq)->sq_ciputctrl; \
616 ASSERT((sq)->sq_type & SQ_CIPUT); \
617 for (i = 0; i <= nlocks; i++) { \
618 ASSERT(MUTEX_HELD(&cip[i].ciputctrl_lock)); \
619 cip[i].ciputctrl_count &= ~SQ_FASTPUT; \
620 } \
621 } \
622 }
623
624 /*
625 * Run service procedures for all queues in the stream head.
626 */
627 #define STR_SERVICE(stp, q) { \
628 ASSERT(MUTEX_HELD(&stp->sd_qlock)); \
629 while (stp->sd_qhead != NULL) { \
630 DQ(q, stp->sd_qhead, stp->sd_qtail, q_link); \
631 ASSERT(stp->sd_nqueues > 0); \
632 stp->sd_nqueues--; \
633 ASSERT(!(q->q_flag & QINSERVICE)); \
634 mutex_exit(&stp->sd_qlock); \
635 queue_service(q); \
636 mutex_enter(&stp->sd_qlock); \
637 } \
638 ASSERT(stp->sd_nqueues == 0); \
639 ASSERT((stp->sd_qhead == NULL) && (stp->sd_qtail == NULL)); \
640 }
641
642 /*
643 * Constructor/destructor routines for the stream head cache
644 */
645 /* ARGSUSED */
646 static int
647 stream_head_constructor(void *buf, void *cdrarg, int kmflags)
648 {
649 stdata_t *stp = buf;
650
651 mutex_init(&stp->sd_lock, NULL, MUTEX_DEFAULT, NULL);
652 mutex_init(&stp->sd_reflock, NULL, MUTEX_DEFAULT, NULL);
653 mutex_init(&stp->sd_qlock, NULL, MUTEX_DEFAULT, NULL);
654 mutex_init(&stp->sd_pid_list_lock, NULL, MUTEX_DEFAULT, NULL);
655 cv_init(&stp->sd_monitor, NULL, CV_DEFAULT, NULL);
656 cv_init(&stp->sd_iocmonitor, NULL, CV_DEFAULT, NULL);
657 cv_init(&stp->sd_refmonitor, NULL, CV_DEFAULT, NULL);
658 cv_init(&stp->sd_qcv, NULL, CV_DEFAULT, NULL);
659 cv_init(&stp->sd_zcopy_wait, NULL, CV_DEFAULT, NULL);
660 list_create(&stp->sd_pid_list, sizeof (pid_node_t),
661 offsetof(pid_node_t, pn_ref_link));
662 stp->sd_wrq = NULL;
663
664 return (0);
665 }
666
667 /* ARGSUSED */
668 static void
669 stream_head_destructor(void *buf, void *cdrarg)
670 {
671 stdata_t *stp = buf;
672
673 mutex_destroy(&stp->sd_lock);
674 mutex_destroy(&stp->sd_reflock);
675 mutex_destroy(&stp->sd_qlock);
676 mutex_destroy(&stp->sd_pid_list_lock);
677 cv_destroy(&stp->sd_monitor);
678 cv_destroy(&stp->sd_iocmonitor);
679 cv_destroy(&stp->sd_refmonitor);
680 cv_destroy(&stp->sd_qcv);
681 cv_destroy(&stp->sd_zcopy_wait);
682 list_destroy(&stp->sd_pid_list);
683 }
684
685 /*
686 * Constructor/destructor routines for the queue cache
687 */
688 /* ARGSUSED */
689 static int
690 queue_constructor(void *buf, void *cdrarg, int kmflags)
691 {
692 queinfo_t *qip = buf;
693 queue_t *qp = &qip->qu_rqueue;
694 queue_t *wqp = &qip->qu_wqueue;
695 syncq_t *sq = &qip->qu_syncq;
696
697 qp->q_first = NULL;
698 qp->q_link = NULL;
699 qp->q_count = 0;
700 qp->q_mblkcnt = 0;
701 qp->q_sqhead = NULL;
702 qp->q_sqtail = NULL;
703 qp->q_sqnext = NULL;
704 qp->q_sqprev = NULL;
705 qp->q_sqflags = 0;
706 qp->q_rwcnt = 0;
707 qp->q_spri = 0;
708
709 mutex_init(QLOCK(qp), NULL, MUTEX_DEFAULT, NULL);
710 cv_init(&qp->q_wait, NULL, CV_DEFAULT, NULL);
711
712 wqp->q_first = NULL;
713 wqp->q_link = NULL;
714 wqp->q_count = 0;
715 wqp->q_mblkcnt = 0;
716 wqp->q_sqhead = NULL;
717 wqp->q_sqtail = NULL;
718 wqp->q_sqnext = NULL;
719 wqp->q_sqprev = NULL;
720 wqp->q_sqflags = 0;
721 wqp->q_rwcnt = 0;
722 wqp->q_spri = 0;
723
724 mutex_init(QLOCK(wqp), NULL, MUTEX_DEFAULT, NULL);
725 cv_init(&wqp->q_wait, NULL, CV_DEFAULT, NULL);
726
727 sq->sq_head = NULL;
728 sq->sq_tail = NULL;
729 sq->sq_evhead = NULL;
730 sq->sq_evtail = NULL;
731 sq->sq_callbpend = NULL;
732 sq->sq_outer = NULL;
733 sq->sq_onext = NULL;
734 sq->sq_oprev = NULL;
735 sq->sq_next = NULL;
736 sq->sq_svcflags = 0;
737 sq->sq_servcount = 0;
738 sq->sq_needexcl = 0;
739 sq->sq_nqueues = 0;
740 sq->sq_pri = 0;
741
742 mutex_init(&sq->sq_lock, NULL, MUTEX_DEFAULT, NULL);
743 cv_init(&sq->sq_wait, NULL, CV_DEFAULT, NULL);
744 cv_init(&sq->sq_exitwait, NULL, CV_DEFAULT, NULL);
745
746 return (0);
747 }
748
749 /* ARGSUSED */
750 static void
751 queue_destructor(void *buf, void *cdrarg)
752 {
753 queinfo_t *qip = buf;
754 queue_t *qp = &qip->qu_rqueue;
755 queue_t *wqp = &qip->qu_wqueue;
756 syncq_t *sq = &qip->qu_syncq;
757
758 ASSERT(qp->q_sqhead == NULL);
759 ASSERT(wqp->q_sqhead == NULL);
760 ASSERT(qp->q_sqnext == NULL);
761 ASSERT(wqp->q_sqnext == NULL);
762 ASSERT(qp->q_rwcnt == 0);
763 ASSERT(wqp->q_rwcnt == 0);
764
765 mutex_destroy(&qp->q_lock);
766 cv_destroy(&qp->q_wait);
767
768 mutex_destroy(&wqp->q_lock);
769 cv_destroy(&wqp->q_wait);
770
771 mutex_destroy(&sq->sq_lock);
772 cv_destroy(&sq->sq_wait);
773 cv_destroy(&sq->sq_exitwait);
774 }
775
776 /*
777 * Constructor/destructor routines for the syncq cache
778 */
779 /* ARGSUSED */
780 static int
781 syncq_constructor(void *buf, void *cdrarg, int kmflags)
782 {
783 syncq_t *sq = buf;
784
785 bzero(buf, sizeof (syncq_t));
786
787 mutex_init(&sq->sq_lock, NULL, MUTEX_DEFAULT, NULL);
788 cv_init(&sq->sq_wait, NULL, CV_DEFAULT, NULL);
789 cv_init(&sq->sq_exitwait, NULL, CV_DEFAULT, NULL);
790
791 return (0);
792 }
793
794 /* ARGSUSED */
795 static void
796 syncq_destructor(void *buf, void *cdrarg)
797 {
798 syncq_t *sq = buf;
799
800 ASSERT(sq->sq_head == NULL);
801 ASSERT(sq->sq_tail == NULL);
802 ASSERT(sq->sq_evhead == NULL);
803 ASSERT(sq->sq_evtail == NULL);
804 ASSERT(sq->sq_callbpend == NULL);
805 ASSERT(sq->sq_callbflags == 0);
806 ASSERT(sq->sq_outer == NULL);
807 ASSERT(sq->sq_onext == NULL);
808 ASSERT(sq->sq_oprev == NULL);
809 ASSERT(sq->sq_next == NULL);
810 ASSERT(sq->sq_needexcl == 0);
811 ASSERT(sq->sq_svcflags == 0);
812 ASSERT(sq->sq_servcount == 0);
813 ASSERT(sq->sq_nqueues == 0);
814 ASSERT(sq->sq_pri == 0);
815 ASSERT(sq->sq_count == 0);
816 ASSERT(sq->sq_rmqcount == 0);
817 ASSERT(sq->sq_cancelid == 0);
818 ASSERT(sq->sq_ciputctrl == NULL);
819 ASSERT(sq->sq_nciputctrl == 0);
820 ASSERT(sq->sq_type == 0);
821 ASSERT(sq->sq_flags == 0);
822
823 mutex_destroy(&sq->sq_lock);
824 cv_destroy(&sq->sq_wait);
825 cv_destroy(&sq->sq_exitwait);
826 }
827
828 /* ARGSUSED */
829 static int
830 ciputctrl_constructor(void *buf, void *cdrarg, int kmflags)
831 {
832 ciputctrl_t *cip = buf;
833 int i;
834
835 for (i = 0; i < n_ciputctrl; i++) {
836 cip[i].ciputctrl_count = SQ_FASTPUT;
837 mutex_init(&cip[i].ciputctrl_lock, NULL, MUTEX_DEFAULT, NULL);
838 }
839
840 return (0);
841 }
842
843 /* ARGSUSED */
844 static void
845 ciputctrl_destructor(void *buf, void *cdrarg)
846 {
847 ciputctrl_t *cip = buf;
848 int i;
849
850 for (i = 0; i < n_ciputctrl; i++) {
851 ASSERT(cip[i].ciputctrl_count & SQ_FASTPUT);
852 mutex_destroy(&cip[i].ciputctrl_lock);
853 }
854 }
855
856 /*
857 * Init routine run from main at boot time.
858 */
859 void
860 strinit(void)
861 {
862 int ncpus = ((boot_max_ncpus == -1) ? max_ncpus : boot_max_ncpus);
863
864 stream_head_cache = kmem_cache_create("stream_head_cache",
865 sizeof (stdata_t), 0,
866 stream_head_constructor, stream_head_destructor, NULL,
867 NULL, NULL, 0);
868
869 queue_cache = kmem_cache_create("queue_cache", sizeof (queinfo_t), 0,
870 queue_constructor, queue_destructor, NULL, NULL, NULL, 0);
871
872 syncq_cache = kmem_cache_create("syncq_cache", sizeof (syncq_t), 0,
873 syncq_constructor, syncq_destructor, NULL, NULL, NULL, 0);
874
875 qband_cache = kmem_cache_create("qband_cache",
876 sizeof (qband_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
877
878 linkinfo_cache = kmem_cache_create("linkinfo_cache",
879 sizeof (linkinfo_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
880
881 n_ciputctrl = ncpus;
882 n_ciputctrl = 1 << highbit(n_ciputctrl - 1);
883 ASSERT(n_ciputctrl >= 1);
884 n_ciputctrl = MIN(n_ciputctrl, max_n_ciputctrl);
885 if (n_ciputctrl >= min_n_ciputctrl) {
886 ciputctrl_cache = kmem_cache_create("ciputctrl_cache",
887 sizeof (ciputctrl_t) * n_ciputctrl,
888 sizeof (ciputctrl_t), ciputctrl_constructor,
889 ciputctrl_destructor, NULL, NULL, NULL, 0);
890 }
891
892 streams_taskq = system_taskq;
893
894 if (streams_taskq == NULL)
895 panic("strinit: no memory for streams taskq!");
896
897 bc_bkgrnd_thread = thread_create(NULL, 0,
898 streams_bufcall_service, NULL, 0, &p0, TS_RUN, streams_lopri);
899
900 streams_qbkgrnd_thread = thread_create(NULL, 0,
901 streams_qbkgrnd_service, NULL, 0, &p0, TS_RUN, streams_lopri);
902
903 streams_sqbkgrnd_thread = thread_create(NULL, 0,
904 streams_sqbkgrnd_service, NULL, 0, &p0, TS_RUN, streams_lopri);
905
906 /*
907 * Create STREAMS kstats.
908 */
909 str_kstat = kstat_create("streams", 0, "strstat",
910 "net", KSTAT_TYPE_NAMED,
911 sizeof (str_statistics) / sizeof (kstat_named_t),
912 KSTAT_FLAG_VIRTUAL);
913
914 if (str_kstat != NULL) {
915 str_kstat->ks_data = &str_statistics;
916 kstat_install(str_kstat);
917 }
918
919 /*
920 * TPI support routine initialisation.
921 */
922 tpi_init();
923
924 /*
925 * Handle to have autopush and persistent link information per
926 * zone.
927 * Note: uses shutdown hook instead of destroy hook so that the
928 * persistent links can be torn down before the destroy hooks
929 * in the TCP/IP stack are called.
930 */
931 netstack_register(NS_STR, str_stack_init, str_stack_shutdown,
932 str_stack_fini);
933 }
934
935 void
936 str_sendsig(vnode_t *vp, int event, uchar_t band, int error)
937 {
938 struct stdata *stp;
939
940 ASSERT(vp->v_stream);
941 stp = vp->v_stream;
942 /* Have to hold sd_lock to prevent siglist from changing */
943 mutex_enter(&stp->sd_lock);
944 if (stp->sd_sigflags & event)
945 strsendsig(stp->sd_siglist, event, band, error);
946 mutex_exit(&stp->sd_lock);
947 }
948
949 /*
950 * Send the "sevent" set of signals to a process.
951 * This might send more than one signal if the process is registered
952 * for multiple events. The caller should pass in an sevent that only
953 * includes the events for which the process has registered.
954 */
955 static void
956 dosendsig(proc_t *proc, int events, int sevent, k_siginfo_t *info,
957 uchar_t band, int error)
958 {
959 ASSERT(MUTEX_HELD(&proc->p_lock));
960
961 info->si_band = 0;
962 info->si_errno = 0;
963
964 if (sevent & S_ERROR) {
965 sevent &= ~S_ERROR;
966 info->si_code = POLL_ERR;
967 info->si_errno = error;
968 TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
969 "strsendsig:proc %p info %p", proc, info);
970 sigaddq(proc, NULL, info, KM_NOSLEEP);
971 info->si_errno = 0;
972 }
973 if (sevent & S_HANGUP) {
974 sevent &= ~S_HANGUP;
975 info->si_code = POLL_HUP;
976 TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
977 "strsendsig:proc %p info %p", proc, info);
978 sigaddq(proc, NULL, info, KM_NOSLEEP);
979 }
980 if (sevent & S_HIPRI) {
981 sevent &= ~S_HIPRI;
982 info->si_code = POLL_PRI;
983 TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
984 "strsendsig:proc %p info %p", proc, info);
985 sigaddq(proc, NULL, info, KM_NOSLEEP);
986 }
987 if (sevent & S_RDBAND) {
988 sevent &= ~S_RDBAND;
989 if (events & S_BANDURG)
990 sigtoproc(proc, NULL, SIGURG);
991 else
992 sigtoproc(proc, NULL, SIGPOLL);
993 }
994 if (sevent & S_WRBAND) {
995 sevent &= ~S_WRBAND;
996 sigtoproc(proc, NULL, SIGPOLL);
997 }
998 if (sevent & S_INPUT) {
999 sevent &= ~S_INPUT;
1000 info->si_code = POLL_IN;
1001 info->si_band = band;
1002 TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
1003 "strsendsig:proc %p info %p", proc, info);
1004 sigaddq(proc, NULL, info, KM_NOSLEEP);
1005 info->si_band = 0;
1006 }
1007 if (sevent & S_OUTPUT) {
1008 sevent &= ~S_OUTPUT;
1009 info->si_code = POLL_OUT;
1010 info->si_band = band;
1011 TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
1012 "strsendsig:proc %p info %p", proc, info);
1013 sigaddq(proc, NULL, info, KM_NOSLEEP);
1014 info->si_band = 0;
1015 }
1016 if (sevent & S_MSG) {
1017 sevent &= ~S_MSG;
1018 info->si_code = POLL_MSG;
1019 info->si_band = band;
1020 TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
1021 "strsendsig:proc %p info %p", proc, info);
1022 sigaddq(proc, NULL, info, KM_NOSLEEP);
1023 info->si_band = 0;
1024 }
1025 if (sevent & S_RDNORM) {
1026 sevent &= ~S_RDNORM;
1027 sigtoproc(proc, NULL, SIGPOLL);
1028 }
1029 if (sevent != 0) {
1030 panic("strsendsig: unknown event(s) %x", sevent);
1031 }
1032 }
1033
1034 /*
1035 * Send SIGPOLL/SIGURG signal to all processes and process groups
1036 * registered on the given signal list that want a signal for at
1037 * least one of the specified events.
1038 *
1039 * Must be called with exclusive access to siglist (caller holding sd_lock).
1040 *
1041 * strioctl(I_SETSIG/I_ESETSIG) will only change siglist when holding
1042 * sd_lock and the ioctl code maintains a PID_HOLD on the pid structure
1043 * while it is in the siglist.
1044 *
1045 * For performance reasons (MP scalability) the code drops pidlock
1046 * when sending signals to a single process.
1047 * When sending to a process group the code holds
1048 * pidlock to prevent the membership in the process group from changing
1049 * while walking the p_pglink list.
1050 */
1051 void
1052 strsendsig(strsig_t *siglist, int event, uchar_t band, int error)
1053 {
1054 strsig_t *ssp;
1055 k_siginfo_t info;
1056 struct pid *pidp;
1057 proc_t *proc;
1058
1059 info.si_signo = SIGPOLL;
1060 info.si_errno = 0;
1061 for (ssp = siglist; ssp; ssp = ssp->ss_next) {
1062 int sevent;
1063
1064 sevent = ssp->ss_events & event;
1065 if (sevent == 0)
1066 continue;
1067
1068 if ((pidp = ssp->ss_pidp) == NULL) {
1069 /* pid was released but still on event list */
1070 continue;
1071 }
1072
1073
1074 if (ssp->ss_pid > 0) {
1075 /*
1076 * XXX This unfortunately still generates
1077 * a signal when a fd is closed but
1078 * the proc is active.
1079 */
1080 ASSERT(ssp->ss_pid == pidp->pid_id);
1081
1082 mutex_enter(&pidlock);
1083 proc = prfind_zone(pidp->pid_id, ALL_ZONES);
1084 if (proc == NULL) {
1085 mutex_exit(&pidlock);
1086 continue;
1087 }
1088 mutex_enter(&proc->p_lock);
1089 mutex_exit(&pidlock);
1090 dosendsig(proc, ssp->ss_events, sevent, &info,
1091 band, error);
1092 mutex_exit(&proc->p_lock);
1093 } else {
1094 /*
1095 * Send to process group. Hold pidlock across
1096 * calls to dosendsig().
1097 */
1098 pid_t pgrp = -ssp->ss_pid;
1099
1100 mutex_enter(&pidlock);
1101 proc = pgfind_zone(pgrp, ALL_ZONES);
1102 while (proc != NULL) {
1103 mutex_enter(&proc->p_lock);
1104 dosendsig(proc, ssp->ss_events, sevent,
1105 &info, band, error);
1106 mutex_exit(&proc->p_lock);
1107 proc = proc->p_pglink;
1108 }
1109 mutex_exit(&pidlock);
1110 }
1111 }
1112 }
1113
1114 /*
1115 * Attach a stream device or module.
1116 * qp is a read queue; the new queue goes in so its next
1117 * read ptr is the argument, and the write queue corresponding
1118 * to the argument points to this queue. Return 0 on success,
1119 * or a non-zero errno on failure.
1120 */
1121 int
1122 qattach(queue_t *qp, dev_t *devp, int oflag, cred_t *crp, fmodsw_impl_t *fp,
1123 boolean_t is_insert)
1124 {
1125 major_t major;
1126 cdevsw_impl_t *dp;
1127 struct streamtab *str;
1128 queue_t *rq;
1129 queue_t *wrq;
1130 uint32_t qflag;
1131 uint32_t sqtype;
1132 perdm_t *dmp;
1133 int error;
1134 int sflag;
1135
1136 rq = allocq();
1137 wrq = _WR(rq);
1138 STREAM(rq) = STREAM(wrq) = STREAM(qp);
1139
1140 if (fp != NULL) {
1141 str = fp->f_str;
1142 qflag = fp->f_qflag;
1143 sqtype = fp->f_sqtype;
1144 dmp = fp->f_dmp;
1145 IMPLY((qflag & (QPERMOD | QMTOUTPERIM)), dmp != NULL);
1146 sflag = MODOPEN;
1147
1148 /*
1149 * stash away a pointer to the module structure so we can
1150 * unref it in qdetach.
1151 */
1152 rq->q_fp = fp;
1153 } else {
1154 ASSERT(!is_insert);
1155
1156 major = getmajor(*devp);
1157 dp = &devimpl[major];
1158
1159 str = dp->d_str;
1160 ASSERT(str == STREAMSTAB(major));
1161
1162 qflag = dp->d_qflag;
1163 ASSERT(qflag & QISDRV);
1164 sqtype = dp->d_sqtype;
1165
1166 /* create perdm_t if needed */
1167 if (NEED_DM(dp->d_dmp, qflag))
1168 dp->d_dmp = hold_dm(str, qflag, sqtype);
1169
1170 dmp = dp->d_dmp;
1171 sflag = 0;
1172 }
1173
1174 TRACE_2(TR_FAC_STREAMS_FR, TR_QATTACH_FLAGS,
1175 "qattach:qflag == %X(%X)", qflag, *devp);
1176
1177 /* setq might sleep in allocator - avoid holding locks. */
1178 setq(rq, str->st_rdinit, str->st_wrinit, dmp, qflag, sqtype, B_FALSE);
1179
1180 /*
1181 * Before calling the module's open routine, set up the q_next
1182 * pointer for inserting a module in the middle of a stream.
1183 *
1184 * Note that we can always set _QINSERTING and set up q_next
1185 * pointer for both inserting and pushing a module. Then there
1186 * is no need for the is_insert parameter. In insertq(), called
1187 * by qprocson(), assume that q_next of the new module always points
1188 * to the correct queue and use it for insertion. Everything should
1189 * work out fine. But in the first release of _I_INSERT, we
1190 * distinguish between inserting and pushing to make sure that
1191 * pushing a module follows the same code path as before.
1192 */
1193 if (is_insert) {
1194 rq->q_flag |= _QINSERTING;
1195 rq->q_next = qp;
1196 }
1197
1198 /*
1199 * If there is an outer perimeter get exclusive access during
1200 * the open procedure. Bump up the reference count on the queue.
1201 */
1202 entersq(rq->q_syncq, SQ_OPENCLOSE);
1203 error = (*rq->q_qinfo->qi_qopen)(rq, devp, oflag, sflag, crp);
1204 if (error != 0)
1205 goto failed;
1206 leavesq(rq->q_syncq, SQ_OPENCLOSE);
1207 ASSERT(qprocsareon(rq));
1208 return (0);
1209
1210 failed:
1211 rq->q_flag &= ~_QINSERTING;
1212 if (backq(wrq) != NULL && backq(wrq)->q_next == wrq)
1213 qprocsoff(rq);
1214 leavesq(rq->q_syncq, SQ_OPENCLOSE);
1215 rq->q_next = wrq->q_next = NULL;
1216 qdetach(rq, 0, 0, crp, B_FALSE);
1217 return (error);
1218 }
1219
1220 /*
1221 * Handle second open of stream. For modules, set the
1222 * last argument to MODOPEN and do not pass any open flags.
1223 * Ignore dummydev since this is not the first open.
1224 */
1225 int
1226 qreopen(queue_t *qp, dev_t *devp, int flag, cred_t *crp)
1227 {
1228 int error;
1229 dev_t dummydev;
1230 queue_t *wqp = _WR(qp);
1231
1232 ASSERT(qp->q_flag & QREADR);
1233 entersq(qp->q_syncq, SQ_OPENCLOSE);
1234
1235 dummydev = *devp;
1236 if (error = ((*qp->q_qinfo->qi_qopen)(qp, &dummydev,
1237 (wqp->q_next ? 0 : flag), (wqp->q_next ? MODOPEN : 0), crp))) {
1238 leavesq(qp->q_syncq, SQ_OPENCLOSE);
1239 mutex_enter(&STREAM(qp)->sd_lock);
1240 qp->q_stream->sd_flag |= STREOPENFAIL;
1241 mutex_exit(&STREAM(qp)->sd_lock);
1242 return (error);
1243 }
1244 leavesq(qp->q_syncq, SQ_OPENCLOSE);
1245
1246 /*
1247 * successful open should have done qprocson()
1248 */
1249 ASSERT(qprocsareon(_RD(qp)));
1250 return (0);
1251 }
1252
1253 /*
1254 * Detach a stream module or device.
1255 * If clmode == 1 then the module or driver was opened and its
1256 * close routine must be called. If clmode == 0, the module
1257 * or driver was never opened or the open failed, and so its close
1258 * should not be called.
1259 */
1260 void
1261 qdetach(queue_t *qp, int clmode, int flag, cred_t *crp, boolean_t is_remove)
1262 {
1263 queue_t *wqp = _WR(qp);
1264 ASSERT(STREAM(qp)->sd_flag & (STRCLOSE|STWOPEN|STRPLUMB));
1265
1266 if (STREAM_NEEDSERVICE(STREAM(qp)))
1267 stream_runservice(STREAM(qp));
1268
1269 if (clmode) {
1270 /*
1271 * Make sure that all the messages on the write side syncq are
1272 * processed and nothing is left. Since we are closing, no new
1273 * messages may appear there.
1274 */
1275 wait_q_syncq(wqp);
1276
1277 entersq(qp->q_syncq, SQ_OPENCLOSE);
1278 if (is_remove) {
1279 mutex_enter(QLOCK(qp));
1280 qp->q_flag |= _QREMOVING;
1281 mutex_exit(QLOCK(qp));
1282 }
1283 (*qp->q_qinfo->qi_qclose)(qp, flag, crp);
1284 /*
1285 * Check that qprocsoff() was actually called.
1286 */
1287 ASSERT((qp->q_flag & QWCLOSE) && (wqp->q_flag & QWCLOSE));
1288
1289 leavesq(qp->q_syncq, SQ_OPENCLOSE);
1290 } else {
1291 disable_svc(qp);
1292 }
1293
1294 /*
1295 * Allow any threads blocked in entersq to proceed and discover
1296 * the QWCLOSE is set.
1297 * Note: This assumes that all users of entersq check QWCLOSE.
1298 * Currently runservice is the only entersq that can happen
1299 * after removeq has finished.
1300 * Removeq will have discarded all messages destined to the closing
1301 * pair of queues from the syncq.
1302 * NOTE: Calling a function inside an assert is unconventional.
1303 * However, it does not cause any problem since flush_syncq() does
1304 * not change any state except when it returns non-zero i.e.
1305 * when the assert will trigger.
1306 */
1307 ASSERT(flush_syncq(qp->q_syncq, qp) == 0);
1308 ASSERT(flush_syncq(wqp->q_syncq, wqp) == 0);
1309 ASSERT((qp->q_flag & QPERMOD) ||
1310 ((qp->q_syncq->sq_head == NULL) &&
1311 (wqp->q_syncq->sq_head == NULL)));
1312
1313 /* release any fmodsw_impl_t structure held on behalf of the queue */
1314 ASSERT(qp->q_fp != NULL || qp->q_flag & QISDRV);
1315 if (qp->q_fp != NULL)
1316 fmodsw_rele(qp->q_fp);
1317
1318 /* freeq removes us from the outer perimeter if any */
1319 freeq(qp);
1320 }
1321
1322 /* Prevent service procedures from being called */
1323 void
1324 disable_svc(queue_t *qp)
1325 {
1326 queue_t *wqp = _WR(qp);
1327
1328 ASSERT(qp->q_flag & QREADR);
1329 mutex_enter(QLOCK(qp));
1330 qp->q_flag |= QWCLOSE;
1331 mutex_exit(QLOCK(qp));
1332 mutex_enter(QLOCK(wqp));
1333 wqp->q_flag |= QWCLOSE;
1334 mutex_exit(QLOCK(wqp));
1335 }
1336
1337 /* Allow service procedures to be called again */
1338 void
1339 enable_svc(queue_t *qp)
1340 {
1341 queue_t *wqp = _WR(qp);
1342
1343 ASSERT(qp->q_flag & QREADR);
1344 mutex_enter(QLOCK(qp));
1345 qp->q_flag &= ~QWCLOSE;
1346 mutex_exit(QLOCK(qp));
1347 mutex_enter(QLOCK(wqp));
1348 wqp->q_flag &= ~QWCLOSE;
1349 mutex_exit(QLOCK(wqp));
1350 }
1351
1352 /*
1353 * Remove queue from qhead/qtail if it is enabled.
1354 * Only reset QENAB if the queue was removed from the runlist.
1355 * A queue goes through 3 stages:
1356 * It is on the service list and QENAB is set.
1357 * It is removed from the service list but QENAB is still set.
1358 * QENAB gets changed to QINSERVICE.
1359 * QINSERVICE is reset (when the service procedure is done)
1360 * Thus we can not reset QENAB unless we actually removed it from the service
1361 * queue.
1362 */
1363 void
1364 remove_runlist(queue_t *qp)
1365 {
1366 if (qp->q_flag & QENAB && qhead != NULL) {
1367 queue_t *q_chase;
1368 queue_t *q_curr;
1369 int removed;
1370
1371 mutex_enter(&service_queue);
1372 RMQ(qp, qhead, qtail, q_link, q_chase, q_curr, removed);
1373 mutex_exit(&service_queue);
1374 if (removed) {
1375 STRSTAT(qremoved);
1376 qp->q_flag &= ~QENAB;
1377 }
1378 }
1379 }
1380
1381
1382 /*
1383 * Wait for any pending service processing to complete.
1384 * The removal of queues from the runlist is not atomic with the
1385 * clearing of the QENABLED flag and setting the INSERVICE flag.
1386 * consequently it is possible for remove_runlist in strclose
1387 * to not find the queue on the runlist but for it to be QENABLED
1388 * and not yet INSERVICE -> hence wait_svc needs to check QENABLED
1389 * as well as INSERVICE.
1390 */
1391 void
1392 wait_svc(queue_t *qp)
1393 {
1394 queue_t *wqp = _WR(qp);
1395
1396 ASSERT(qp->q_flag & QREADR);
1397
1398 /*
1399 * Try to remove queues from qhead/qtail list.
1400 */
1401 if (qhead != NULL) {
1402 remove_runlist(qp);
1403 remove_runlist(wqp);
1404 }
1405 /*
1406 * Wait till the syncqs associated with the queue disappear from the
1407 * background processing list.
1408 * This only needs to be done for non-PERMOD perimeters since
1409 * for PERMOD perimeters the syncq may be shared and will only be freed
1410 * when the last module/driver is unloaded.
1411 * If for PERMOD perimeters queue was on the syncq list, removeq()
1412 * should call propagate_syncq() or drain_syncq() for it. Both of these
1413 * functions remove the queue from its syncq list, so sqthread will not
1414 * try to access the queue.
1415 */
1416 if (!(qp->q_flag & QPERMOD)) {
1417 syncq_t *rsq = qp->q_syncq;
1418 syncq_t *wsq = wqp->q_syncq;
1419
1420 /*
1421 * Disable rsq and wsq and wait for any background processing of
1422 * syncq to complete.
1423 */
1424 wait_sq_svc(rsq);
1425 if (wsq != rsq)
1426 wait_sq_svc(wsq);
1427 }
1428
1429 mutex_enter(QLOCK(qp));
1430 while (qp->q_flag & (QINSERVICE|QENAB))
1431 cv_wait(&qp->q_wait, QLOCK(qp));
1432 mutex_exit(QLOCK(qp));
1433 mutex_enter(QLOCK(wqp));
1434 while (wqp->q_flag & (QINSERVICE|QENAB))
1435 cv_wait(&wqp->q_wait, QLOCK(wqp));
1436 mutex_exit(QLOCK(wqp));
1437 }
1438
1439 /*
1440 * Put ioctl data from userland buffer `arg' into the mblk chain `bp'.
1441 * `flag' must always contain either K_TO_K or U_TO_K; STR_NOSIG may
1442 * also be set, and is passed through to allocb_cred_wait().
1443 *
1444 * Returns errno on failure, zero on success.
1445 */
1446 int
1447 putiocd(mblk_t *bp, char *arg, int flag, cred_t *cr)
1448 {
1449 mblk_t *tmp;
1450 ssize_t count;
1451 int error = 0;
1452
1453 ASSERT((flag & (U_TO_K | K_TO_K)) == U_TO_K ||
1454 (flag & (U_TO_K | K_TO_K)) == K_TO_K);
1455
1456 if (bp->b_datap->db_type == M_IOCTL) {
1457 count = ((struct iocblk *)bp->b_rptr)->ioc_count;
1458 } else {
1459 ASSERT(bp->b_datap->db_type == M_COPYIN);
1460 count = ((struct copyreq *)bp->b_rptr)->cq_size;
1461 }
1462 /*
1463 * strdoioctl validates ioc_count, so if this assert fails it
1464 * cannot be due to user error.
1465 */
1466 ASSERT(count >= 0);
1467
1468 if ((tmp = allocb_cred_wait(count, (flag & STR_NOSIG), &error, cr,
1469 curproc->p_pid)) == NULL) {
1470 return (error);
1471 }
1472 error = strcopyin(arg, tmp->b_wptr, count, flag & (U_TO_K|K_TO_K));
1473 if (error != 0) {
1474 freeb(tmp);
1475 return (error);
1476 }
1477 DB_CPID(tmp) = curproc->p_pid;
1478 tmp->b_wptr += count;
1479 bp->b_cont = tmp;
1480
1481 return (0);
1482 }
1483
1484 /*
1485 * Copy ioctl data to user-land. Return non-zero errno on failure,
1486 * 0 for success.
1487 */
1488 int
1489 getiocd(mblk_t *bp, char *arg, int copymode)
1490 {
1491 ssize_t count;
1492 size_t n;
1493 int error;
1494
1495 if (bp->b_datap->db_type == M_IOCACK)
1496 count = ((struct iocblk *)bp->b_rptr)->ioc_count;
1497 else {
1498 ASSERT(bp->b_datap->db_type == M_COPYOUT);
1499 count = ((struct copyreq *)bp->b_rptr)->cq_size;
1500 }
1501 ASSERT(count >= 0);
1502
1503 for (bp = bp->b_cont; bp && count;
1504 count -= n, bp = bp->b_cont, arg += n) {
1505 n = MIN(count, bp->b_wptr - bp->b_rptr);
1506 error = strcopyout(bp->b_rptr, arg, n, copymode);
1507 if (error)
1508 return (error);
1509 }
1510 ASSERT(count == 0);
1511 return (0);
1512 }
1513
1514 /*
1515 * Allocate a linkinfo entry given the write queue of the
1516 * bottom module of the top stream and the write queue of the
1517 * stream head of the bottom stream.
1518 */
1519 linkinfo_t *
1520 alloclink(queue_t *qup, queue_t *qdown, file_t *fpdown)
1521 {
1522 linkinfo_t *linkp;
1523
1524 linkp = kmem_cache_alloc(linkinfo_cache, KM_SLEEP);
1525
1526 linkp->li_lblk.l_qtop = qup;
1527 linkp->li_lblk.l_qbot = qdown;
1528 linkp->li_fpdown = fpdown;
1529
1530 mutex_enter(&strresources);
1531 linkp->li_next = linkinfo_list;
1532 linkp->li_prev = NULL;
1533 if (linkp->li_next)
1534 linkp->li_next->li_prev = linkp;
1535 linkinfo_list = linkp;
1536 linkp->li_lblk.l_index = ++lnk_id;
1537 ASSERT(lnk_id != 0); /* this should never wrap in practice */
1538 mutex_exit(&strresources);
1539
1540 return (linkp);
1541 }
1542
1543 /*
1544 * Free a linkinfo entry.
1545 */
1546 void
1547 lbfree(linkinfo_t *linkp)
1548 {
1549 mutex_enter(&strresources);
1550 if (linkp->li_next)
1551 linkp->li_next->li_prev = linkp->li_prev;
1552 if (linkp->li_prev)
1553 linkp->li_prev->li_next = linkp->li_next;
1554 else
1555 linkinfo_list = linkp->li_next;
1556 mutex_exit(&strresources);
1557
1558 kmem_cache_free(linkinfo_cache, linkp);
1559 }
1560
1561 /*
1562 * Check for a potential linking cycle.
1563 * Return 1 if a link will result in a cycle,
1564 * and 0 otherwise.
1565 */
1566 int
1567 linkcycle(stdata_t *upstp, stdata_t *lostp, str_stack_t *ss)
1568 {
1569 struct mux_node *np;
1570 struct mux_edge *ep;
1571 int i;
1572 major_t lomaj;
1573 major_t upmaj;
1574 /*
1575 * if the lower stream is a pipe/FIFO, return, since link
1576 * cycles can not happen on pipes/FIFOs
1577 */
1578 if (lostp->sd_vnode->v_type == VFIFO)
1579 return (0);
1580
1581 for (i = 0; i < ss->ss_devcnt; i++) {
1582 np = &ss->ss_mux_nodes[i];
1583 MUX_CLEAR(np);
1584 }
1585 lomaj = getmajor(lostp->sd_vnode->v_rdev);
1586 upmaj = getmajor(upstp->sd_vnode->v_rdev);
1587 np = &ss->ss_mux_nodes[lomaj];
1588 for (;;) {
1589 if (!MUX_DIDVISIT(np)) {
1590 if (np->mn_imaj == upmaj)
1591 return (1);
1592 if (np->mn_outp == NULL) {
1593 MUX_VISIT(np);
1594 if (np->mn_originp == NULL)
1595 return (0);
1596 np = np->mn_originp;
1597 continue;
1598 }
1599 MUX_VISIT(np);
1600 np->mn_startp = np->mn_outp;
1601 } else {
1602 if (np->mn_startp == NULL) {
1603 if (np->mn_originp == NULL)
1604 return (0);
1605 else {
1606 np = np->mn_originp;
1607 continue;
1608 }
1609 }
1610 /*
1611 * If ep->me_nodep is a FIFO (me_nodep == NULL),
1612 * ignore the edge and move on. ep->me_nodep gets
1613 * set to NULL in mux_addedge() if it is a FIFO.
1614 *
1615 */
1616 ep = np->mn_startp;
1617 np->mn_startp = ep->me_nextp;
1618 if (ep->me_nodep == NULL)
1619 continue;
1620 ep->me_nodep->mn_originp = np;
1621 np = ep->me_nodep;
1622 }
1623 }
1624 }
1625
1626 /*
1627 * Find linkinfo entry corresponding to the parameters.
1628 */
1629 linkinfo_t *
1630 findlinks(stdata_t *stp, int index, int type, str_stack_t *ss)
1631 {
1632 linkinfo_t *linkp;
1633 struct mux_edge *mep;
1634 struct mux_node *mnp;
1635 queue_t *qup;
1636
1637 mutex_enter(&strresources);
1638 if ((type & LINKTYPEMASK) == LINKNORMAL) {
1639 qup = getendq(stp->sd_wrq);
1640 for (linkp = linkinfo_list; linkp; linkp = linkp->li_next) {
1641 if ((qup == linkp->li_lblk.l_qtop) &&
1642 (!index || (index == linkp->li_lblk.l_index))) {
1643 mutex_exit(&strresources);
1644 return (linkp);
1645 }
1646 }
1647 } else {
1648 ASSERT((type & LINKTYPEMASK) == LINKPERSIST);
1649 mnp = &ss->ss_mux_nodes[getmajor(stp->sd_vnode->v_rdev)];
1650 mep = mnp->mn_outp;
1651 while (mep) {
1652 if ((index == 0) || (index == mep->me_muxid))
1653 break;
1654 mep = mep->me_nextp;
1655 }
1656 if (!mep) {
1657 mutex_exit(&strresources);
1658 return (NULL);
1659 }
1660 for (linkp = linkinfo_list; linkp; linkp = linkp->li_next) {
1661 if ((!linkp->li_lblk.l_qtop) &&
1662 (mep->me_muxid == linkp->li_lblk.l_index)) {
1663 mutex_exit(&strresources);
1664 return (linkp);
1665 }
1666 }
1667 }
1668 mutex_exit(&strresources);
1669 return (NULL);
1670 }
1671
1672 /*
1673 * Given a queue ptr, follow the chain of q_next pointers until you reach the
1674 * last queue on the chain and return it.
1675 */
1676 queue_t *
1677 getendq(queue_t *q)
1678 {
1679 ASSERT(q != NULL);
1680 while (_SAMESTR(q))
1681 q = q->q_next;
1682 return (q);
1683 }
1684
1685 /*
1686 * Wait for the syncq count to drop to zero.
1687 * sq could be either outer or inner.
1688 */
1689
1690 static void
1691 wait_syncq(syncq_t *sq)
1692 {
1693 uint16_t count;
1694
1695 mutex_enter(SQLOCK(sq));
1696 count = sq->sq_count;
1697 SQ_PUTLOCKS_ENTER(sq);
1698 SUM_SQ_PUTCOUNTS(sq, count);
1699 while (count != 0) {
1700 sq->sq_flags |= SQ_WANTWAKEUP;
1701 SQ_PUTLOCKS_EXIT(sq);
1702 cv_wait(&sq->sq_wait, SQLOCK(sq));
1703 count = sq->sq_count;
1704 SQ_PUTLOCKS_ENTER(sq);
1705 SUM_SQ_PUTCOUNTS(sq, count);
1706 }
1707 SQ_PUTLOCKS_EXIT(sq);
1708 mutex_exit(SQLOCK(sq));
1709 }
1710
1711 /*
1712 * Wait while there are any messages for the queue in its syncq.
1713 */
1714 static void
1715 wait_q_syncq(queue_t *q)
1716 {
1717 if ((q->q_sqflags & Q_SQQUEUED) || (q->q_syncqmsgs > 0)) {
1718 syncq_t *sq = q->q_syncq;
1719
1720 mutex_enter(SQLOCK(sq));
1721 while ((q->q_sqflags & Q_SQQUEUED) || (q->q_syncqmsgs > 0)) {
1722 sq->sq_flags |= SQ_WANTWAKEUP;
1723 cv_wait(&sq->sq_wait, SQLOCK(sq));
1724 }
1725 mutex_exit(SQLOCK(sq));
1726 }
1727 }
1728
1729
1730 int
1731 mlink_file(vnode_t *vp, int cmd, struct file *fpdown, cred_t *crp, int *rvalp,
1732 int lhlink)
1733 {
1734 struct stdata *stp;
1735 struct strioctl strioc;
1736 struct linkinfo *linkp;
1737 struct stdata *stpdown;
1738 struct streamtab *str;
1739 queue_t *passq;
1740 syncq_t *passyncq;
1741 queue_t *rq;
1742 cdevsw_impl_t *dp;
1743 uint32_t qflag;
1744 uint32_t sqtype;
1745 perdm_t *dmp;
1746 int error = 0;
1747 netstack_t *ns;
1748 str_stack_t *ss;
1749
1750 stp = vp->v_stream;
1751 TRACE_1(TR_FAC_STREAMS_FR,
1752 TR_I_LINK, "I_LINK/I_PLINK:stp %p", stp);
1753 /*
1754 * Test for invalid upper stream
1755 */
1756 if (stp->sd_flag & STRHUP) {
1757 return (ENXIO);
1758 }
1759 if (vp->v_type == VFIFO) {
1760 return (EINVAL);
1761 }
1762 if (stp->sd_strtab == NULL) {
1763 return (EINVAL);
1764 }
1765 if (!stp->sd_strtab->st_muxwinit) {
1766 return (EINVAL);
1767 }
1768 if (fpdown == NULL) {
1769 return (EBADF);
1770 }
1771 ns = netstack_find_by_cred(crp);
1772 ASSERT(ns != NULL);
1773 ss = ns->netstack_str;
1774 ASSERT(ss != NULL);
1775
1776 if (getmajor(stp->sd_vnode->v_rdev) >= ss->ss_devcnt) {
1777 netstack_rele(ss->ss_netstack);
1778 return (EINVAL);
1779 }
1780 mutex_enter(&muxifier);
1781 if (stp->sd_flag & STPLEX) {
1782 mutex_exit(&muxifier);
1783 netstack_rele(ss->ss_netstack);
1784 return (ENXIO);
1785 }
1786
1787 /*
1788 * Test for invalid lower stream.
1789 * The check for the v_type != VFIFO and having a major
1790 * number not >= devcnt is done to avoid problems with
1791 * adding mux_node entry past the end of mux_nodes[].
1792 * For FIFO's we don't add an entry so this isn't a
1793 * problem.
1794 */
1795 if (((stpdown = fpdown->f_vnode->v_stream) == NULL) ||
1796 (stpdown == stp) || (stpdown->sd_flag &
1797 (STPLEX|STRHUP|STRDERR|STWRERR|IOCWAIT|STRPLUMB)) ||
1798 ((stpdown->sd_vnode->v_type != VFIFO) &&
1799 (getmajor(stpdown->sd_vnode->v_rdev) >= ss->ss_devcnt)) ||
1800 linkcycle(stp, stpdown, ss)) {
1801 mutex_exit(&muxifier);
1802 netstack_rele(ss->ss_netstack);
1803 return (EINVAL);
1804 }
1805 TRACE_1(TR_FAC_STREAMS_FR,
1806 TR_STPDOWN, "stpdown:%p", stpdown);
1807 rq = getendq(stp->sd_wrq);
1808 if (cmd == I_PLINK)
1809 rq = NULL;
1810
1811 linkp = alloclink(rq, stpdown->sd_wrq, fpdown);
1812
1813 strioc.ic_cmd = cmd;
1814 strioc.ic_timout = INFTIM;
1815 strioc.ic_len = sizeof (struct linkblk);
1816 strioc.ic_dp = (char *)&linkp->li_lblk;
1817
1818 /*
1819 * STRPLUMB protects plumbing changes and should be set before
1820 * link_addpassthru()/link_rempassthru() are called, so it is set here
1821 * and cleared in the end of mlink when passthru queue is removed.
1822 * Setting of STRPLUMB prevents reopens of the stream while passthru
1823 * queue is in-place (it is not a proper module and doesn't have open
1824 * entry point).
1825 *
1826 * STPLEX prevents any threads from entering the stream from above. It
1827 * can't be set before the call to link_addpassthru() because putnext
1828 * from below may cause stream head I/O routines to be called and these
1829 * routines assert that STPLEX is not set. After link_addpassthru()
1830 * nothing may come from below since the pass queue syncq is blocked.
1831 * Note also that STPLEX should be cleared before the call to
1832 * link_rempassthru() since when messages start flowing to the stream
1833 * head (e.g. because of message propagation from the pass queue) stream
1834 * head I/O routines may be called with STPLEX flag set.
1835 *
1836 * When STPLEX is set, nothing may come into the stream from above and
1837 * it is safe to do a setq which will change stream head. So, the
1838 * correct sequence of actions is:
1839 *
1840 * 1) Set STRPLUMB
1841 * 2) Call link_addpassthru()
1842 * 3) Set STPLEX
1843 * 4) Call setq and update the stream state
1844 * 5) Clear STPLEX
1845 * 6) Call link_rempassthru()
1846 * 7) Clear STRPLUMB
1847 *
1848 * The same sequence applies to munlink() code.
1849 */
1850 mutex_enter(&stpdown->sd_lock);
1851 stpdown->sd_flag |= STRPLUMB;
1852 mutex_exit(&stpdown->sd_lock);
1853 /*
1854 * Add passthru queue below lower mux. This will block
1855 * syncqs of lower muxs read queue during I_LINK/I_UNLINK.
1856 */
1857 passq = link_addpassthru(stpdown);
1858
1859 mutex_enter(&stpdown->sd_lock);
1860 stpdown->sd_flag |= STPLEX;
1861 mutex_exit(&stpdown->sd_lock);
1862
1863 rq = _RD(stpdown->sd_wrq);
1864 /*
1865 * There may be messages in the streamhead's syncq due to messages
1866 * that arrived before link_addpassthru() was done. To avoid
1867 * background processing of the syncq happening simultaneous with
1868 * setq processing, we disable the streamhead syncq and wait until
1869 * existing background thread finishes working on it.
1870 */
1871 wait_sq_svc(rq->q_syncq);
1872 passyncq = passq->q_syncq;
1873 if (!(passyncq->sq_flags & SQ_BLOCKED))
1874 blocksq(passyncq, SQ_BLOCKED, 0);
1875
1876 ASSERT((rq->q_flag & QMT_TYPEMASK) == QMTSAFE);
1877 ASSERT(rq->q_syncq == SQ(rq) && _WR(rq)->q_syncq == SQ(rq));
1878 rq->q_ptr = _WR(rq)->q_ptr = NULL;
1879
1880 /* setq might sleep in allocator - avoid holding locks. */
1881 /* Note: we are holding muxifier here. */
1882
1883 str = stp->sd_strtab;
1884 dp = &devimpl[getmajor(vp->v_rdev)];
1885 ASSERT(dp->d_str == str);
1886
1887 qflag = dp->d_qflag;
1888 sqtype = dp->d_sqtype;
1889
1890 /* create perdm_t if needed */
1891 if (NEED_DM(dp->d_dmp, qflag))
1892 dp->d_dmp = hold_dm(str, qflag, sqtype);
1893
1894 dmp = dp->d_dmp;
1895
1896 setq(rq, str->st_muxrinit, str->st_muxwinit, dmp, qflag, sqtype,
1897 B_TRUE);
1898
1899 /*
1900 * XXX Remove any "odd" messages from the queue.
1901 * Keep only M_DATA, M_PROTO, M_PCPROTO.
1902 */
1903 error = strdoioctl(stp, &strioc, FNATIVE,
1904 K_TO_K | STR_NOERROR | STR_NOSIG, crp, rvalp);
1905 if (error != 0) {
1906 lbfree(linkp);
1907
1908 if (!(passyncq->sq_flags & SQ_BLOCKED))
1909 blocksq(passyncq, SQ_BLOCKED, 0);
1910 /*
1911 * Restore the stream head queue and then remove
1912 * the passq. Turn off STPLEX before we turn on
1913 * the stream by removing the passq.
1914 */
1915 rq->q_ptr = _WR(rq)->q_ptr = stpdown;
1916 setq(rq, &strdata, &stwdata, NULL, QMTSAFE, SQ_CI|SQ_CO,
1917 B_TRUE);
1918
1919 mutex_enter(&stpdown->sd_lock);
1920 stpdown->sd_flag &= ~STPLEX;
1921 mutex_exit(&stpdown->sd_lock);
1922
1923 link_rempassthru(passq);
1924
1925 mutex_enter(&stpdown->sd_lock);
1926 stpdown->sd_flag &= ~STRPLUMB;
1927 /* Wakeup anyone waiting for STRPLUMB to clear. */
1928 cv_broadcast(&stpdown->sd_monitor);
1929 mutex_exit(&stpdown->sd_lock);
1930
1931 mutex_exit(&muxifier);
1932 netstack_rele(ss->ss_netstack);
1933 return (error);
1934 }
1935 mutex_enter(&fpdown->f_tlock);
1936 fpdown->f_count++;
1937 mutex_exit(&fpdown->f_tlock);
1938
1939 /*
1940 * if we've made it here the linkage is all set up so we should also
1941 * set up the layered driver linkages
1942 */
1943
1944 ASSERT((cmd == I_LINK) || (cmd == I_PLINK));
1945 if (cmd == I_LINK) {
1946 ldi_mlink_fp(stp, fpdown, lhlink, LINKNORMAL);
1947 } else {
1948 ldi_mlink_fp(stp, fpdown, lhlink, LINKPERSIST);
1949 }
1950
1951 link_rempassthru(passq);
1952
1953 mux_addedge(stp, stpdown, linkp->li_lblk.l_index, ss);
1954
1955 /*
1956 * Mark the upper stream as having dependent links
1957 * so that strclose can clean it up.
1958 */
1959 if (cmd == I_LINK) {
1960 mutex_enter(&stp->sd_lock);
1961 stp->sd_flag |= STRHASLINKS;
1962 mutex_exit(&stp->sd_lock);
1963 }
1964 /*
1965 * Wake up any other processes that may have been
1966 * waiting on the lower stream. These will all
1967 * error out.
1968 */
1969 mutex_enter(&stpdown->sd_lock);
1970 /* The passthru module is removed so we may release STRPLUMB */
1971 stpdown->sd_flag &= ~STRPLUMB;
1972 cv_broadcast(&rq->q_wait);
1973 cv_broadcast(&_WR(rq)->q_wait);
1974 cv_broadcast(&stpdown->sd_monitor);
1975 mutex_exit(&stpdown->sd_lock);
1976 mutex_exit(&muxifier);
1977 *rvalp = linkp->li_lblk.l_index;
1978 netstack_rele(ss->ss_netstack);
1979 return (0);
1980 }
1981
1982 int
1983 mlink(vnode_t *vp, int cmd, int arg, cred_t *crp, int *rvalp, int lhlink)
1984 {
1985 int ret;
1986 struct file *fpdown;
1987
1988 fpdown = getf(arg);
1989 ret = mlink_file(vp, cmd, fpdown, crp, rvalp, lhlink);
1990 if (fpdown != NULL)
1991 releasef(arg);
1992 return (ret);
1993 }
1994
1995 /*
1996 * Unlink a multiplexor link. Stp is the controlling stream for the
1997 * link, and linkp points to the link's entry in the linkinfo list.
1998 * The muxifier lock must be held on entry and is dropped on exit.
1999 *
2000 * NOTE : Currently it is assumed that mux would process all the messages
2001 * sitting on it's queue before ACKing the UNLINK. It is the responsibility
2002 * of the mux to handle all the messages that arrive before UNLINK.
2003 * If the mux has to send down messages on its lower stream before
2004 * ACKing I_UNLINK, then it *should* know to handle messages even
2005 * after the UNLINK is acked (actually it should be able to handle till we
2006 * re-block the read side of the pass queue here). If the mux does not
2007 * open up the lower stream, any messages that arrive during UNLINK
2008 * will be put in the stream head. In the case of lower stream opening
2009 * up, some messages might land in the stream head depending on when
2010 * the message arrived and when the read side of the pass queue was
2011 * re-blocked.
2012 */
2013 int
2014 munlink(stdata_t *stp, linkinfo_t *linkp, int flag, cred_t *crp, int *rvalp,
2015 str_stack_t *ss)
2016 {
2017 struct strioctl strioc;
2018 struct stdata *stpdown;
2019 queue_t *rq, *wrq;
2020 queue_t *passq;
2021 syncq_t *passyncq;
2022 int error = 0;
2023 file_t *fpdown;
2024
2025 ASSERT(MUTEX_HELD(&muxifier));
2026
2027 stpdown = linkp->li_fpdown->f_vnode->v_stream;
2028
2029 /*
2030 * See the comment in mlink() concerning STRPLUMB/STPLEX flags.
2031 */
2032 mutex_enter(&stpdown->sd_lock);
2033 stpdown->sd_flag |= STRPLUMB;
2034 mutex_exit(&stpdown->sd_lock);
2035
2036 /*
2037 * Add passthru queue below lower mux. This will block
2038 * syncqs of lower muxs read queue during I_LINK/I_UNLINK.
2039 */
2040 passq = link_addpassthru(stpdown);
2041
2042 if ((flag & LINKTYPEMASK) == LINKNORMAL)
2043 strioc.ic_cmd = I_UNLINK;
2044 else
2045 strioc.ic_cmd = I_PUNLINK;
2046 strioc.ic_timout = INFTIM;
2047 strioc.ic_len = sizeof (struct linkblk);
2048 strioc.ic_dp = (char *)&linkp->li_lblk;
2049
2050 error = strdoioctl(stp, &strioc, FNATIVE,
2051 K_TO_K | STR_NOERROR | STR_NOSIG, crp, rvalp);
2052
2053 /*
2054 * If there was an error and this is not called via strclose,
2055 * return to the user. Otherwise, pretend there was no error
2056 * and close the link.
2057 */
2058 if (error) {
2059 if (flag & LINKCLOSE) {
2060 cmn_err(CE_WARN, "KERNEL: munlink: could not perform "
2061 "unlink ioctl, closing anyway (%d)\n", error);
2062 } else {
2063 link_rempassthru(passq);
2064 mutex_enter(&stpdown->sd_lock);
2065 stpdown->sd_flag &= ~STRPLUMB;
2066 cv_broadcast(&stpdown->sd_monitor);
2067 mutex_exit(&stpdown->sd_lock);
2068 mutex_exit(&muxifier);
2069 return (error);
2070 }
2071 }
2072
2073 mux_rmvedge(stp, linkp->li_lblk.l_index, ss);
2074 fpdown = linkp->li_fpdown;
2075 lbfree(linkp);
2076
2077 /*
2078 * We go ahead and drop muxifier here--it's a nasty global lock that
2079 * can slow others down. It's okay to since attempts to mlink() this
2080 * stream will be stopped because STPLEX is still set in the stdata
2081 * structure, and munlink() is stopped because mux_rmvedge() and
2082 * lbfree() have removed it from mux_nodes[] and linkinfo_list,
2083 * respectively. Note that we defer the closef() of fpdown until
2084 * after we drop muxifier since strclose() can call munlinkall().
2085 */
2086 mutex_exit(&muxifier);
2087
2088 wrq = stpdown->sd_wrq;
2089 rq = _RD(wrq);
2090
2091 /*
2092 * Get rid of outstanding service procedure runs, before we make
2093 * it a stream head, since a stream head doesn't have any service
2094 * procedure.
2095 */
2096 disable_svc(rq);
2097 wait_svc(rq);
2098
2099 /*
2100 * Since we don't disable the syncq for QPERMOD, we wait for whatever
2101 * is queued up to be finished. mux should take care that nothing is
2102 * send down to this queue. We should do it now as we're going to block
2103 * passyncq if it was unblocked.
2104 */
2105 if (wrq->q_flag & QPERMOD) {
2106 syncq_t *sq = wrq->q_syncq;
2107
2108 mutex_enter(SQLOCK(sq));
2109 while (wrq->q_sqflags & Q_SQQUEUED) {
2110 sq->sq_flags |= SQ_WANTWAKEUP;
2111 cv_wait(&sq->sq_wait, SQLOCK(sq));
2112 }
2113 mutex_exit(SQLOCK(sq));
2114 }
2115 passyncq = passq->q_syncq;
2116 if (!(passyncq->sq_flags & SQ_BLOCKED)) {
2117
2118 syncq_t *sq, *outer;
2119
2120 /*
2121 * Messages could be flowing from underneath. We will
2122 * block the read side of the passq. This would be
2123 * sufficient for QPAIR and QPERQ muxes to ensure
2124 * that no data is flowing up into this queue
2125 * and hence no thread active in this instance of
2126 * lower mux. But for QPERMOD and QMTOUTPERIM there
2127 * could be messages on the inner and outer/inner
2128 * syncqs respectively. We will wait for them to drain.
2129 * Because passq is blocked messages end up in the syncq
2130 * And qfill_syncq could possibly end up setting QFULL
2131 * which will access the rq->q_flag. Hence, we have to
2132 * acquire the QLOCK in setq.
2133 *
2134 * XXX Messages can also flow from top into this
2135 * queue though the unlink is over (Ex. some instance
2136 * in putnext() called from top that has still not
2137 * accessed this queue. And also putq(lowerq) ?).
2138 * Solution : How about blocking the l_qtop queue ?
2139 * Do we really care about such pure D_MP muxes ?
2140 */
2141
2142 blocksq(passyncq, SQ_BLOCKED, 0);
2143
2144 sq = rq->q_syncq;
2145 if ((outer = sq->sq_outer) != NULL) {
2146
2147 /*
2148 * We have to just wait for the outer sq_count
2149 * drop to zero. As this does not prevent new
2150 * messages to enter the outer perimeter, this
2151 * is subject to starvation.
2152 *
2153 * NOTE :Because of blocksq above, messages could
2154 * be in the inner syncq only because of some
2155 * thread holding the outer perimeter exclusively.
2156 * Hence it would be sufficient to wait for the
2157 * exclusive holder of the outer perimeter to drain
2158 * the inner and outer syncqs. But we will not depend
2159 * on this feature and hence check the inner syncqs
2160 * separately.
2161 */
2162 wait_syncq(outer);
2163 }
2164
2165
2166 /*
2167 * There could be messages destined for
2168 * this queue. Let the exclusive holder
2169 * drain it.
2170 */
2171
2172 wait_syncq(sq);
2173 ASSERT((rq->q_flag & QPERMOD) ||
2174 ((rq->q_syncq->sq_head == NULL) &&
2175 (_WR(rq)->q_syncq->sq_head == NULL)));
2176 }
2177
2178 /*
2179 * We haven't taken care of QPERMOD case yet. QPERMOD is a special
2180 * case as we don't disable its syncq or remove it off the syncq
2181 * service list.
2182 */
2183 if (rq->q_flag & QPERMOD) {
2184 syncq_t *sq = rq->q_syncq;
2185
2186 mutex_enter(SQLOCK(sq));
2187 while (rq->q_sqflags & Q_SQQUEUED) {
2188 sq->sq_flags |= SQ_WANTWAKEUP;
2189 cv_wait(&sq->sq_wait, SQLOCK(sq));
2190 }
2191 mutex_exit(SQLOCK(sq));
2192 }
2193
2194 /*
2195 * flush_syncq changes states only when there are some messages to
2196 * free, i.e. when it returns non-zero value to return.
2197 */
2198 ASSERT(flush_syncq(rq->q_syncq, rq) == 0);
2199 ASSERT(flush_syncq(wrq->q_syncq, wrq) == 0);
2200
2201 /*
2202 * Nobody else should know about this queue now.
2203 * If the mux did not process the messages before
2204 * acking the I_UNLINK, free them now.
2205 */
2206
2207 flushq(rq, FLUSHALL);
2208 flushq(_WR(rq), FLUSHALL);
2209
2210 /*
2211 * Convert the mux lower queue into a stream head queue.
2212 * Turn off STPLEX before we turn on the stream by removing the passq.
2213 */
2214 rq->q_ptr = wrq->q_ptr = stpdown;
2215 setq(rq, &strdata, &stwdata, NULL, QMTSAFE, SQ_CI|SQ_CO, B_TRUE);
2216
2217 ASSERT((rq->q_flag & QMT_TYPEMASK) == QMTSAFE);
2218 ASSERT(rq->q_syncq == SQ(rq) && _WR(rq)->q_syncq == SQ(rq));
2219
2220 enable_svc(rq);
2221
2222 /*
2223 * Now it is a proper stream, so STPLEX is cleared. But STRPLUMB still
2224 * needs to be set to prevent reopen() of the stream - such reopen may
2225 * try to call non-existent pass queue open routine and panic.
2226 */
2227 mutex_enter(&stpdown->sd_lock);
2228 stpdown->sd_flag &= ~STPLEX;
2229 mutex_exit(&stpdown->sd_lock);
2230
2231 ASSERT(((flag & LINKTYPEMASK) == LINKNORMAL) ||
2232 ((flag & LINKTYPEMASK) == LINKPERSIST));
2233
2234 /* clean up the layered driver linkages */
2235 if ((flag & LINKTYPEMASK) == LINKNORMAL) {
2236 ldi_munlink_fp(stp, fpdown, LINKNORMAL);
2237 } else {
2238 ldi_munlink_fp(stp, fpdown, LINKPERSIST);
2239 }
2240
2241 link_rempassthru(passq);
2242
2243 /*
2244 * Now all plumbing changes are finished and STRPLUMB is no
2245 * longer needed.
2246 */
2247 mutex_enter(&stpdown->sd_lock);
2248 stpdown->sd_flag &= ~STRPLUMB;
2249 cv_broadcast(&stpdown->sd_monitor);
2250 mutex_exit(&stpdown->sd_lock);
2251
2252 (void) closef(fpdown);
2253 return (0);
2254 }
2255
2256 /*
2257 * Unlink all multiplexor links for which stp is the controlling stream.
2258 * Return 0, or a non-zero errno on failure.
2259 */
2260 int
2261 munlinkall(stdata_t *stp, int flag, cred_t *crp, int *rvalp, str_stack_t *ss)
2262 {
2263 linkinfo_t *linkp;
2264 int error = 0;
2265
2266 mutex_enter(&muxifier);
2267 while (linkp = findlinks(stp, 0, flag, ss)) {
2268 /*
2269 * munlink() releases the muxifier lock.
2270 */
2271 if (error = munlink(stp, linkp, flag, crp, rvalp, ss))
2272 return (error);
2273 mutex_enter(&muxifier);
2274 }
2275 mutex_exit(&muxifier);
2276 return (0);
2277 }
2278
2279 /*
2280 * A multiplexor link has been made. Add an
2281 * edge to the directed graph.
2282 */
2283 void
2284 mux_addedge(stdata_t *upstp, stdata_t *lostp, int muxid, str_stack_t *ss)
2285 {
2286 struct mux_node *np;
2287 struct mux_edge *ep;
2288 major_t upmaj;
2289 major_t lomaj;
2290
2291 upmaj = getmajor(upstp->sd_vnode->v_rdev);
2292 lomaj = getmajor(lostp->sd_vnode->v_rdev);
2293 np = &ss->ss_mux_nodes[upmaj];
2294 if (np->mn_outp) {
2295 ep = np->mn_outp;
2296 while (ep->me_nextp)
2297 ep = ep->me_nextp;
2298 ep->me_nextp = kmem_alloc(sizeof (struct mux_edge), KM_SLEEP);
2299 ep = ep->me_nextp;
2300 } else {
2301 np->mn_outp = kmem_alloc(sizeof (struct mux_edge), KM_SLEEP);
2302 ep = np->mn_outp;
2303 }
2304 ep->me_nextp = NULL;
2305 ep->me_muxid = muxid;
2306 /*
2307 * Save the dev_t for the purposes of str_stack_shutdown.
2308 * str_stack_shutdown assumes that the device allows reopen, since
2309 * this dev_t is the one after any cloning by xx_open().
2310 * Would prefer finding the dev_t from before any cloning,
2311 * but specfs doesn't retain that.
2312 */
2313 ep->me_dev = upstp->sd_vnode->v_rdev;
2314 if (lostp->sd_vnode->v_type == VFIFO)
2315 ep->me_nodep = NULL;
2316 else
2317 ep->me_nodep = &ss->ss_mux_nodes[lomaj];
2318 }
2319
2320 /*
2321 * A multiplexor link has been removed. Remove the
2322 * edge in the directed graph.
2323 */
2324 void
2325 mux_rmvedge(stdata_t *upstp, int muxid, str_stack_t *ss)
2326 {
2327 struct mux_node *np;
2328 struct mux_edge *ep;
2329 struct mux_edge *pep = NULL;
2330 major_t upmaj;
2331
2332 upmaj = getmajor(upstp->sd_vnode->v_rdev);
2333 np = &ss->ss_mux_nodes[upmaj];
2334 ASSERT(np->mn_outp != NULL);
2335 ep = np->mn_outp;
2336 while (ep) {
2337 if (ep->me_muxid == muxid) {
2338 if (pep)
2339 pep->me_nextp = ep->me_nextp;
2340 else
2341 np->mn_outp = ep->me_nextp;
2342 kmem_free(ep, sizeof (struct mux_edge));
2343 return;
2344 }
2345 pep = ep;
2346 ep = ep->me_nextp;
2347 }
2348 ASSERT(0); /* should not reach here */
2349 }
2350
2351 /*
2352 * Translate the device flags (from conf.h) to the corresponding
2353 * qflag and sq_flag (type) values.
2354 */
2355 int
2356 devflg_to_qflag(struct streamtab *stp, uint32_t devflag, uint32_t *qflagp,
2357 uint32_t *sqtypep)
2358 {
2359 uint32_t qflag = 0;
2360 uint32_t sqtype = 0;
2361
2362 if (devflag & _D_OLD)
2363 goto bad;
2364
2365 /* Inner perimeter presence and scope */
2366 switch (devflag & D_MTINNER_MASK) {
2367 case D_MP:
2368 qflag |= QMTSAFE;
2369 sqtype |= SQ_CI;
2370 break;
2371 case D_MTPERQ|D_MP:
2372 qflag |= QPERQ;
2373 break;
2374 case D_MTQPAIR|D_MP:
2375 qflag |= QPAIR;
2376 break;
2377 case D_MTPERMOD|D_MP:
2378 qflag |= QPERMOD;
2379 break;
2380 default:
2381 goto bad;
2382 }
2383
2384 /* Outer perimeter */
2385 if (devflag & D_MTOUTPERIM) {
2386 switch (devflag & D_MTINNER_MASK) {
2387 case D_MP:
2388 case D_MTPERQ|D_MP:
2389 case D_MTQPAIR|D_MP:
2390 break;
2391 default:
2392 goto bad;
2393 }
2394 qflag |= QMTOUTPERIM;
2395 }
2396
2397 /* Inner perimeter modifiers */
2398 if (devflag & D_MTINNER_MOD) {
2399 switch (devflag & D_MTINNER_MASK) {
2400 case D_MP:
2401 goto bad;
2402 default:
2403 break;
2404 }
2405 if (devflag & D_MTPUTSHARED)
2406 sqtype |= SQ_CIPUT;
2407 if (devflag & _D_MTOCSHARED) {
2408 /*
2409 * The code in putnext assumes that it has the
2410 * highest concurrency by not checking sq_count.
2411 * Thus _D_MTOCSHARED can only be supported when
2412 * D_MTPUTSHARED is set.
2413 */
2414 if (!(devflag & D_MTPUTSHARED))
2415 goto bad;
2416 sqtype |= SQ_CIOC;
2417 }
2418 if (devflag & _D_MTCBSHARED) {
2419 /*
2420 * The code in putnext assumes that it has the
2421 * highest concurrency by not checking sq_count.
2422 * Thus _D_MTCBSHARED can only be supported when
2423 * D_MTPUTSHARED is set.
2424 */
2425 if (!(devflag & D_MTPUTSHARED))
2426 goto bad;
2427 sqtype |= SQ_CICB;
2428 }
2429 if (devflag & _D_MTSVCSHARED) {
2430 /*
2431 * The code in putnext assumes that it has the
2432 * highest concurrency by not checking sq_count.
2433 * Thus _D_MTSVCSHARED can only be supported when
2434 * D_MTPUTSHARED is set. Also _D_MTSVCSHARED is
2435 * supported only for QPERMOD.
2436 */
2437 if (!(devflag & D_MTPUTSHARED) || !(qflag & QPERMOD))
2438 goto bad;
2439 sqtype |= SQ_CISVC;
2440 }
2441 }
2442
2443 /* Default outer perimeter concurrency */
2444 sqtype |= SQ_CO;
2445
2446 /* Outer perimeter modifiers */
2447 if (devflag & D_MTOCEXCL) {
2448 if (!(devflag & D_MTOUTPERIM)) {
2449 /* No outer perimeter */
2450 goto bad;
2451 }
2452 sqtype &= ~SQ_COOC;
2453 }
2454
2455 /* Synchronous Streams extended qinit structure */
2456 if (devflag & D_SYNCSTR)
2457 qflag |= QSYNCSTR;
2458
2459 /*
2460 * Private flag used by a transport module to indicate
2461 * to sockfs that it supports direct-access mode without
2462 * having to go through STREAMS.
2463 */
2464 if (devflag & _D_DIRECT) {
2465 /* Reject unless the module is fully-MT (no perimeter) */
2466 if ((qflag & QMT_TYPEMASK) != QMTSAFE)
2467 goto bad;
2468 qflag |= _QDIRECT;
2469 }
2470
2471 *qflagp = qflag;
2472 *sqtypep = sqtype;
2473 return (0);
2474
2475 bad:
2476 cmn_err(CE_WARN,
2477 "stropen: bad MT flags (0x%x) in driver '%s'",
2478 (int)(qflag & D_MTSAFETY_MASK),
2479 stp->st_rdinit->qi_minfo->mi_idname);
2480
2481 return (EINVAL);
2482 }
2483
2484 /*
2485 * Set the interface values for a pair of queues (qinit structure,
2486 * packet sizes, water marks).
2487 * setq assumes that the caller does not have a claim (entersq or claimq)
2488 * on the queue.
2489 */
2490 void
2491 setq(queue_t *rq, struct qinit *rinit, struct qinit *winit,
2492 perdm_t *dmp, uint32_t qflag, uint32_t sqtype, boolean_t lock_needed)
2493 {
2494 queue_t *wq;
2495 syncq_t *sq, *outer;
2496
2497 ASSERT(rq->q_flag & QREADR);
2498 ASSERT((qflag & QMT_TYPEMASK) != 0);
2499 IMPLY((qflag & (QPERMOD | QMTOUTPERIM)), dmp != NULL);
2500
2501 wq = _WR(rq);
2502 rq->q_qinfo = rinit;
2503 rq->q_hiwat = rinit->qi_minfo->mi_hiwat;
2504 rq->q_lowat = rinit->qi_minfo->mi_lowat;
2505 rq->q_minpsz = rinit->qi_minfo->mi_minpsz;
2506 rq->q_maxpsz = rinit->qi_minfo->mi_maxpsz;
2507 wq->q_qinfo = winit;
2508 wq->q_hiwat = winit->qi_minfo->mi_hiwat;
2509 wq->q_lowat = winit->qi_minfo->mi_lowat;
2510 wq->q_minpsz = winit->qi_minfo->mi_minpsz;
2511 wq->q_maxpsz = winit->qi_minfo->mi_maxpsz;
2512
2513 /* Remove old syncqs */
2514 sq = rq->q_syncq;
2515 outer = sq->sq_outer;
2516 if (outer != NULL) {
2517 ASSERT(wq->q_syncq->sq_outer == outer);
2518 outer_remove(outer, rq->q_syncq);
2519 if (wq->q_syncq != rq->q_syncq)
2520 outer_remove(outer, wq->q_syncq);
2521 }
2522 ASSERT(sq->sq_outer == NULL);
2523 ASSERT(sq->sq_onext == NULL && sq->sq_oprev == NULL);
2524
2525 if (sq != SQ(rq)) {
2526 if (!(rq->q_flag & QPERMOD))
2527 free_syncq(sq);
2528 if (wq->q_syncq == rq->q_syncq)
2529 wq->q_syncq = NULL;
2530 rq->q_syncq = NULL;
2531 }
2532 if (wq->q_syncq != NULL && wq->q_syncq != sq &&
2533 wq->q_syncq != SQ(rq)) {
2534 free_syncq(wq->q_syncq);
2535 wq->q_syncq = NULL;
2536 }
2537 ASSERT(rq->q_syncq == NULL || (rq->q_syncq->sq_head == NULL &&
2538 rq->q_syncq->sq_tail == NULL));
2539 ASSERT(wq->q_syncq == NULL || (wq->q_syncq->sq_head == NULL &&
2540 wq->q_syncq->sq_tail == NULL));
2541
2542 if (!(rq->q_flag & QPERMOD) &&
2543 rq->q_syncq != NULL && rq->q_syncq->sq_ciputctrl != NULL) {
2544 ASSERT(rq->q_syncq->sq_nciputctrl == n_ciputctrl - 1);
2545 SUMCHECK_CIPUTCTRL_COUNTS(rq->q_syncq->sq_ciputctrl,
2546 rq->q_syncq->sq_nciputctrl, 0);
2547 ASSERT(ciputctrl_cache != NULL);
2548 kmem_cache_free(ciputctrl_cache, rq->q_syncq->sq_ciputctrl);
2549 rq->q_syncq->sq_ciputctrl = NULL;
2550 rq->q_syncq->sq_nciputctrl = 0;
2551 }
2552
2553 if (!(wq->q_flag & QPERMOD) &&
2554 wq->q_syncq != NULL && wq->q_syncq->sq_ciputctrl != NULL) {
2555 ASSERT(wq->q_syncq->sq_nciputctrl == n_ciputctrl - 1);
2556 SUMCHECK_CIPUTCTRL_COUNTS(wq->q_syncq->sq_ciputctrl,
2557 wq->q_syncq->sq_nciputctrl, 0);
2558 ASSERT(ciputctrl_cache != NULL);
2559 kmem_cache_free(ciputctrl_cache, wq->q_syncq->sq_ciputctrl);
2560 wq->q_syncq->sq_ciputctrl = NULL;
2561 wq->q_syncq->sq_nciputctrl = 0;
2562 }
2563
2564 sq = SQ(rq);
2565 ASSERT(sq->sq_head == NULL && sq->sq_tail == NULL);
2566 ASSERT(sq->sq_outer == NULL);
2567 ASSERT(sq->sq_onext == NULL && sq->sq_oprev == NULL);
2568
2569 /*
2570 * Create syncqs based on qflag and sqtype. Set the SQ_TYPES_IN_FLAGS
2571 * bits in sq_flag based on the sqtype.
2572 */
2573 ASSERT((sq->sq_flags & ~SQ_TYPES_IN_FLAGS) == 0);
2574
2575 rq->q_syncq = wq->q_syncq = sq;
2576 sq->sq_type = sqtype;
2577 sq->sq_flags = (sqtype & SQ_TYPES_IN_FLAGS);
2578
2579 /*
2580 * We are making sq_svcflags zero,
2581 * resetting SQ_DISABLED in case it was set by
2582 * wait_svc() in the munlink path.
2583 *
2584 */
2585 ASSERT((sq->sq_svcflags & SQ_SERVICE) == 0);
2586 sq->sq_svcflags = 0;
2587
2588 /*
2589 * We need to acquire the lock here for the mlink and munlink case,
2590 * where canputnext, backenable, etc can access the q_flag.
2591 */
2592 if (lock_needed) {
2593 mutex_enter(QLOCK(rq));
2594 rq->q_flag = (rq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2595 mutex_exit(QLOCK(rq));
2596 mutex_enter(QLOCK(wq));
2597 wq->q_flag = (wq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2598 mutex_exit(QLOCK(wq));
2599 } else {
2600 rq->q_flag = (rq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2601 wq->q_flag = (wq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2602 }
2603
2604 if (qflag & QPERQ) {
2605 /* Allocate a separate syncq for the write side */
2606 sq = new_syncq();
2607 sq->sq_type = rq->q_syncq->sq_type;
2608 sq->sq_flags = rq->q_syncq->sq_flags;
2609 ASSERT(sq->sq_outer == NULL && sq->sq_onext == NULL &&
2610 sq->sq_oprev == NULL);
2611 wq->q_syncq = sq;
2612 }
2613 if (qflag & QPERMOD) {
2614 sq = dmp->dm_sq;
2615
2616 /*
2617 * Assert that we do have an inner perimeter syncq and that it
2618 * does not have an outer perimeter associated with it.
2619 */
2620 ASSERT(sq->sq_outer == NULL && sq->sq_onext == NULL &&
2621 sq->sq_oprev == NULL);
2622 rq->q_syncq = wq->q_syncq = sq;
2623 }
2624 if (qflag & QMTOUTPERIM) {
2625 outer = dmp->dm_sq;
2626
2627 ASSERT(outer->sq_outer == NULL);
2628 outer_insert(outer, rq->q_syncq);
2629 if (wq->q_syncq != rq->q_syncq)
2630 outer_insert(outer, wq->q_syncq);
2631 }
2632 ASSERT((rq->q_syncq->sq_flags & SQ_TYPES_IN_FLAGS) ==
2633 (rq->q_syncq->sq_type & SQ_TYPES_IN_FLAGS));
2634 ASSERT((wq->q_syncq->sq_flags & SQ_TYPES_IN_FLAGS) ==
2635 (wq->q_syncq->sq_type & SQ_TYPES_IN_FLAGS));
2636 ASSERT((rq->q_flag & QMT_TYPEMASK) == (qflag & QMT_TYPEMASK));
2637
2638 /*
2639 * Initialize struio() types.
2640 */
2641 rq->q_struiot =
2642 (rq->q_flag & QSYNCSTR) ? rinit->qi_struiot : STRUIOT_NONE;
2643 wq->q_struiot =
2644 (wq->q_flag & QSYNCSTR) ? winit->qi_struiot : STRUIOT_NONE;
2645 }
2646
2647 perdm_t *
2648 hold_dm(struct streamtab *str, uint32_t qflag, uint32_t sqtype)
2649 {
2650 syncq_t *sq;
2651 perdm_t **pp;
2652 perdm_t *p;
2653 perdm_t *dmp;
2654
2655 ASSERT(str != NULL);
2656 ASSERT(qflag & (QPERMOD | QMTOUTPERIM));
2657
2658 rw_enter(&perdm_rwlock, RW_READER);
2659 for (p = perdm_list; p != NULL; p = p->dm_next) {
2660 if (p->dm_str == str) { /* found one */
2661 atomic_inc_32(&(p->dm_ref));
2662 rw_exit(&perdm_rwlock);
2663 return (p);
2664 }
2665 }
2666 rw_exit(&perdm_rwlock);
2667
2668 sq = new_syncq();
2669 if (qflag & QPERMOD) {
2670 sq->sq_type = sqtype | SQ_PERMOD;
2671 sq->sq_flags = sqtype & SQ_TYPES_IN_FLAGS;
2672 } else {
2673 ASSERT(qflag & QMTOUTPERIM);
2674 sq->sq_onext = sq->sq_oprev = sq;
2675 }
2676
2677 dmp = kmem_alloc(sizeof (perdm_t), KM_SLEEP);
2678 dmp->dm_sq = sq;
2679 dmp->dm_str = str;
2680 dmp->dm_ref = 1;
2681 dmp->dm_next = NULL;
2682
2683 rw_enter(&perdm_rwlock, RW_WRITER);
2684 for (pp = &perdm_list; (p = *pp) != NULL; pp = &(p->dm_next)) {
2685 if (p->dm_str == str) { /* already present */
2686 p->dm_ref++;
2687 rw_exit(&perdm_rwlock);
2688 free_syncq(sq);
2689 kmem_free(dmp, sizeof (perdm_t));
2690 return (p);
2691 }
2692 }
2693
2694 *pp = dmp;
2695 rw_exit(&perdm_rwlock);
2696 return (dmp);
2697 }
2698
2699 void
2700 rele_dm(perdm_t *dmp)
2701 {
2702 perdm_t **pp;
2703 perdm_t *p;
2704
2705 rw_enter(&perdm_rwlock, RW_WRITER);
2706 ASSERT(dmp->dm_ref > 0);
2707
2708 if (--dmp->dm_ref > 0) {
2709 rw_exit(&perdm_rwlock);
2710 return;
2711 }
2712
2713 for (pp = &perdm_list; (p = *pp) != NULL; pp = &(p->dm_next))
2714 if (p == dmp)
2715 break;
2716 ASSERT(p == dmp);
2717 *pp = p->dm_next;
2718 rw_exit(&perdm_rwlock);
2719
2720 /*
2721 * Wait for any background processing that relies on the
2722 * syncq to complete before it is freed.
2723 */
2724 wait_sq_svc(p->dm_sq);
2725 free_syncq(p->dm_sq);
2726 kmem_free(p, sizeof (perdm_t));
2727 }
2728
2729 /*
2730 * Make a protocol message given control and data buffers.
2731 * n.b., this can block; be careful of what locks you hold when calling it.
2732 *
2733 * If sd_maxblk is less than *iosize this routine can fail part way through
2734 * (due to an allocation failure). In this case on return *iosize will contain
2735 * the amount that was consumed. Otherwise *iosize will not be modified
2736 * i.e. it will contain the amount that was consumed.
2737 */
2738 int
2739 strmakemsg(
2740 struct strbuf *mctl,
2741 ssize_t *iosize,
2742 struct uio *uiop,
2743 stdata_t *stp,
2744 int32_t flag,
2745 mblk_t **mpp)
2746 {
2747 mblk_t *mpctl = NULL;
2748 mblk_t *mpdata = NULL;
2749 int error;
2750
2751 ASSERT(uiop != NULL);
2752
2753 *mpp = NULL;
2754 /* Create control part, if any */
2755 if ((mctl != NULL) && (mctl->len >= 0)) {
2756 error = strmakectl(mctl, flag, uiop->uio_fmode, &mpctl);
2757 if (error)
2758 return (error);
2759 }
2760 /* Create data part, if any */
2761 if (*iosize >= 0) {
2762 error = strmakedata(iosize, uiop, stp, flag, &mpdata);
2763 if (error) {
2764 freemsg(mpctl);
2765 return (error);
2766 }
2767 }
2768 if (mpctl != NULL) {
2769 if (mpdata != NULL)
2770 linkb(mpctl, mpdata);
2771 *mpp = mpctl;
2772 } else {
2773 *mpp = mpdata;
2774 }
2775 return (0);
2776 }
2777
2778 /*
2779 * Make the control part of a protocol message given a control buffer.
2780 * n.b., this can block; be careful of what locks you hold when calling it.
2781 */
2782 int
2783 strmakectl(
2784 struct strbuf *mctl,
2785 int32_t flag,
2786 int32_t fflag,
2787 mblk_t **mpp)
2788 {
2789 mblk_t *bp = NULL;
2790 unsigned char msgtype;
2791 int error = 0;
2792 cred_t *cr = CRED();
2793
2794 /* We do not support interrupt threads using the stream head to send */
2795 ASSERT(cr != NULL);
2796
2797 *mpp = NULL;
2798 /*
2799 * Create control part of message, if any.
2800 */
2801 if ((mctl != NULL) && (mctl->len >= 0)) {
2802 caddr_t base;
2803 int ctlcount;
2804 int allocsz;
2805
2806 if (flag & RS_HIPRI)
2807 msgtype = M_PCPROTO;
2808 else
2809 msgtype = M_PROTO;
2810
2811 ctlcount = mctl->len;
2812 base = mctl->buf;
2813
2814 /*
2815 * Give modules a better chance to reuse M_PROTO/M_PCPROTO
2816 * blocks by increasing the size to something more usable.
2817 */
2818 allocsz = MAX(ctlcount, 64);
2819
2820 /*
2821 * Range checking has already been done; simply try
2822 * to allocate a message block for the ctl part.
2823 */
2824 while ((bp = allocb_cred(allocsz, cr,
2825 curproc->p_pid)) == NULL) {
2826 if (fflag & (FNDELAY|FNONBLOCK))
2827 return (EAGAIN);
2828 if (error = strwaitbuf(allocsz, BPRI_MED))
2829 return (error);
2830 }
2831
2832 bp->b_datap->db_type = msgtype;
2833 if (copyin(base, bp->b_wptr, ctlcount)) {
2834 freeb(bp);
2835 return (EFAULT);
2836 }
2837 bp->b_wptr += ctlcount;
2838 }
2839 *mpp = bp;
2840 return (0);
2841 }
2842
2843 /*
2844 * Make a protocol message given data buffers.
2845 * n.b., this can block; be careful of what locks you hold when calling it.
2846 *
2847 * If sd_maxblk is less than *iosize this routine can fail part way through
2848 * (due to an allocation failure). In this case on return *iosize will contain
2849 * the amount that was consumed. Otherwise *iosize will not be modified
2850 * i.e. it will contain the amount that was consumed.
2851 */
2852 int
2853 strmakedata(
2854 ssize_t *iosize,
2855 struct uio *uiop,
2856 stdata_t *stp,
2857 int32_t flag,
2858 mblk_t **mpp)
2859 {
2860 mblk_t *mp = NULL;
2861 mblk_t *bp;
2862 int wroff = (int)stp->sd_wroff;
2863 int tail_len = (int)stp->sd_tail;
2864 int extra = wroff + tail_len;
2865 int error = 0;
2866 ssize_t maxblk;
2867 ssize_t count = *iosize;
2868 cred_t *cr;
2869
2870 *mpp = NULL;
2871 if (count < 0)
2872 return (0);
2873
2874 /* We do not support interrupt threads using the stream head to send */
2875 cr = CRED();
2876 ASSERT(cr != NULL);
2877
2878 maxblk = stp->sd_maxblk;
2879 if (maxblk == INFPSZ)
2880 maxblk = count;
2881
2882 /*
2883 * Create data part of message, if any.
2884 */
2885 do {
2886 ssize_t size;
2887 dblk_t *dp;
2888
2889 ASSERT(uiop);
2890
2891 size = MIN(count, maxblk);
2892
2893 while ((bp = allocb_cred(size + extra, cr,
2894 curproc->p_pid)) == NULL) {
2895 error = EAGAIN;
2896 if ((uiop->uio_fmode & (FNDELAY|FNONBLOCK)) ||
2897 (error = strwaitbuf(size + extra, BPRI_MED)) != 0) {
2898 if (count == *iosize) {
2899 freemsg(mp);
2900 return (error);
2901 } else {
2902 *iosize -= count;
2903 *mpp = mp;
2904 return (0);
2905 }
2906 }
2907 }
2908 dp = bp->b_datap;
2909 dp->db_cpid = curproc->p_pid;
2910 ASSERT(wroff <= dp->db_lim - bp->b_wptr);
2911 bp->b_wptr = bp->b_rptr = bp->b_rptr + wroff;
2912
2913 if (flag & STRUIO_POSTPONE) {
2914 /*
2915 * Setup the stream uio portion of the
2916 * dblk for subsequent use by struioget().
2917 */
2918 dp->db_struioflag = STRUIO_SPEC;
2919 dp->db_cksumstart = 0;
2920 dp->db_cksumstuff = 0;
2921 dp->db_cksumend = size;
2922 *(long long *)dp->db_struioun.data = 0ll;
2923 bp->b_wptr += size;
2924 } else {
2925 if (stp->sd_copyflag & STRCOPYCACHED)
2926 uiop->uio_extflg |= UIO_COPY_CACHED;
2927
2928 if (size != 0) {
2929 error = uiomove(bp->b_wptr, size, UIO_WRITE,
2930 uiop);
2931 if (error != 0) {
2932 freeb(bp);
2933 freemsg(mp);
2934 return (error);
2935 }
2936 }
2937 bp->b_wptr += size;
2938
2939 if (stp->sd_wputdatafunc != NULL) {
2940 mblk_t *newbp;
2941
2942 newbp = (stp->sd_wputdatafunc)(stp->sd_vnode,
2943 bp, NULL, NULL, NULL, NULL);
2944 if (newbp == NULL) {
2945 freeb(bp);
2946 freemsg(mp);
2947 return (ECOMM);
2948 }
2949 bp = newbp;
2950 }
2951 }
2952
2953 count -= size;
2954
2955 if (mp == NULL)
2956 mp = bp;
2957 else
2958 linkb(mp, bp);
2959 } while (count > 0);
2960
2961 *mpp = mp;
2962 return (0);
2963 }
2964
2965 /*
2966 * Wait for a buffer to become available. Return non-zero errno
2967 * if not able to wait, 0 if buffer is probably there.
2968 */
2969 int
2970 strwaitbuf(size_t size, int pri)
2971 {
2972 bufcall_id_t id;
2973
2974 mutex_enter(&bcall_monitor);
2975 if ((id = bufcall(size, pri, (void (*)(void *))cv_broadcast,
2976 &ttoproc(curthread)->p_flag_cv)) == 0) {
2977 mutex_exit(&bcall_monitor);
2978 return (ENOSR);
2979 }
2980 if (!cv_wait_sig(&(ttoproc(curthread)->p_flag_cv), &bcall_monitor)) {
2981 unbufcall(id);
2982 mutex_exit(&bcall_monitor);
2983 return (EINTR);
2984 }
2985 unbufcall(id);
2986 mutex_exit(&bcall_monitor);
2987 return (0);
2988 }
2989
2990 /*
2991 * This function waits for a read or write event to happen on a stream.
2992 * fmode can specify FNDELAY and/or FNONBLOCK.
2993 * The timeout is in ms with -1 meaning infinite.
2994 * The flag values work as follows:
2995 * READWAIT Check for read side errors, send M_READ
2996 * GETWAIT Check for read side errors, no M_READ
2997 * WRITEWAIT Check for write side errors.
2998 * NOINTR Do not return error if nonblocking or timeout.
2999 * STR_NOERROR Ignore all errors except STPLEX.
3000 * STR_NOSIG Ignore/hold signals during the duration of the call.
3001 * STR_PEEK Pass through the strgeterr().
3002 */
3003 int
3004 strwaitq(stdata_t *stp, int flag, ssize_t count, int fmode, clock_t timout,
3005 int *done)
3006 {
3007 int slpflg, errs;
3008 int error;
3009 kcondvar_t *sleepon;
3010 mblk_t *mp;
3011 ssize_t *rd_count;
3012 clock_t rval;
3013
3014 ASSERT(MUTEX_HELD(&stp->sd_lock));
3015 if ((flag & READWAIT) || (flag & GETWAIT)) {
3016 slpflg = RSLEEP;
3017 sleepon = &_RD(stp->sd_wrq)->q_wait;
3018 errs = STRDERR|STPLEX;
3019 } else {
3020 slpflg = WSLEEP;
3021 sleepon = &stp->sd_wrq->q_wait;
3022 errs = STWRERR|STRHUP|STPLEX;
3023 }
3024 if (flag & STR_NOERROR)
3025 errs = STPLEX;
3026
3027 if (stp->sd_wakeq & slpflg) {
3028 /*
3029 * A strwakeq() is pending, no need to sleep.
3030 */
3031 stp->sd_wakeq &= ~slpflg;
3032 *done = 0;
3033 return (0);
3034 }
3035
3036 if (stp->sd_flag & errs) {
3037 /*
3038 * Check for errors before going to sleep since the
3039 * caller might not have checked this while holding
3040 * sd_lock.
3041 */
3042 error = strgeterr(stp, errs, (flag & STR_PEEK));
3043 if (error != 0) {
3044 *done = 1;
3045 return (error);
3046 }
3047 }
3048
3049 /*
3050 * If any module downstream has requested read notification
3051 * by setting SNDMREAD flag using M_SETOPTS, send a message
3052 * down stream.
3053 */
3054 if ((flag & READWAIT) && (stp->sd_flag & SNDMREAD)) {
3055 mutex_exit(&stp->sd_lock);
3056 if (!(mp = allocb_wait(sizeof (ssize_t), BPRI_MED,
3057 (flag & STR_NOSIG), &error))) {
3058 mutex_enter(&stp->sd_lock);
3059 *done = 1;
3060 return (error);
3061 }
3062 mp->b_datap->db_type = M_READ;
3063 rd_count = (ssize_t *)mp->b_wptr;
3064 *rd_count = count;
3065 mp->b_wptr += sizeof (ssize_t);
3066 /*
3067 * Send the number of bytes requested by the
3068 * read as the argument to M_READ.
3069 */
3070 stream_willservice(stp);
3071 putnext(stp->sd_wrq, mp);
3072 stream_runservice(stp);
3073 mutex_enter(&stp->sd_lock);
3074
3075 /*
3076 * If any data arrived due to inline processing
3077 * of putnext(), don't sleep.
3078 */
3079 if (_RD(stp->sd_wrq)->q_first != NULL) {
3080 *done = 0;
3081 return (0);
3082 }
3083 }
3084
3085 if (fmode & (FNDELAY|FNONBLOCK)) {
3086 if (!(flag & NOINTR))
3087 error = EAGAIN;
3088 else
3089 error = 0;
3090 *done = 1;
3091 return (error);
3092 }
3093
3094 stp->sd_flag |= slpflg;
3095 TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_WAIT2,
3096 "strwaitq sleeps (2):%p, %X, %lX, %X, %p",
3097 stp, flag, count, fmode, done);
3098
3099 rval = str_cv_wait(sleepon, &stp->sd_lock, timout, flag & STR_NOSIG);
3100 if (rval > 0) {
3101 /* EMPTY */
3102 TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_WAKE2,
3103 "strwaitq awakes(2):%X, %X, %X, %X, %X",
3104 stp, flag, count, fmode, done);
3105 } else if (rval == 0) {
3106 TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_INTR2,
3107 "strwaitq interrupt #2:%p, %X, %lX, %X, %p",
3108 stp, flag, count, fmode, done);
3109 stp->sd_flag &= ~slpflg;
3110 cv_broadcast(sleepon);
3111 if (!(flag & NOINTR))
3112 error = EINTR;
3113 else
3114 error = 0;
3115 *done = 1;
3116 return (error);
3117 } else {
3118 /* timeout */
3119 TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_TIME,
3120 "strwaitq timeout:%p, %X, %lX, %X, %p",
3121 stp, flag, count, fmode, done);
3122 *done = 1;
3123 if (!(flag & NOINTR))
3124 return (ETIME);
3125 else
3126 return (0);
3127 }
3128 /*
3129 * If the caller implements delayed errors (i.e. queued after data)
3130 * we can not check for errors here since data as well as an
3131 * error might have arrived at the stream head. We return to
3132 * have the caller check the read queue before checking for errors.
3133 */
3134 if ((stp->sd_flag & errs) && !(flag & STR_DELAYERR)) {
3135 error = strgeterr(stp, errs, (flag & STR_PEEK));
3136 if (error != 0) {
3137 *done = 1;
3138 return (error);
3139 }
3140 }
3141 *done = 0;
3142 return (0);
3143 }
3144
3145 /*
3146 * Perform job control discipline access checks.
3147 * Return 0 for success and the errno for failure.
3148 */
3149
3150 #define cantsend(p, t, sig) \
3151 (sigismember(&(p)->p_ignore, sig) || signal_is_blocked((t), sig))
3152
3153 int
3154 straccess(struct stdata *stp, enum jcaccess mode)
3155 {
3156 extern kcondvar_t lbolt_cv; /* XXX: should be in a header file */
3157 kthread_t *t = curthread;
3158 proc_t *p = ttoproc(t);
3159 sess_t *sp;
3160
3161 ASSERT(mutex_owned(&stp->sd_lock));
3162
3163 if (stp->sd_sidp == NULL || stp->sd_vnode->v_type == VFIFO)
3164 return (0);
3165
3166 mutex_enter(&p->p_lock); /* protects p_pgidp */
3167
3168 for (;;) {
3169 mutex_enter(&p->p_splock); /* protects p->p_sessp */
3170 sp = p->p_sessp;
3171 mutex_enter(&sp->s_lock); /* protects sp->* */
3172
3173 /*
3174 * If this is not the calling process's controlling terminal
3175 * or if the calling process is already in the foreground
3176 * then allow access.
3177 */
3178 if (sp->s_dev != stp->sd_vnode->v_rdev ||
3179 p->p_pgidp == stp->sd_pgidp) {
3180 mutex_exit(&sp->s_lock);
3181 mutex_exit(&p->p_splock);
3182 mutex_exit(&p->p_lock);
3183 return (0);
3184 }
3185
3186 /*
3187 * Check to see if controlling terminal has been deallocated.
3188 */
3189 if (sp->s_vp == NULL) {
3190 if (!cantsend(p, t, SIGHUP))
3191 sigtoproc(p, t, SIGHUP);
3192 mutex_exit(&sp->s_lock);
3193 mutex_exit(&p->p_splock);
3194 mutex_exit(&p->p_lock);
3195 return (EIO);
3196 }
3197
3198 mutex_exit(&sp->s_lock);
3199 mutex_exit(&p->p_splock);
3200
3201 if (mode == JCGETP) {
3202 mutex_exit(&p->p_lock);
3203 return (0);
3204 }
3205
3206 if (mode == JCREAD) {
3207 if (p->p_detached || cantsend(p, t, SIGTTIN)) {
3208 mutex_exit(&p->p_lock);
3209 return (EIO);
3210 }
3211 mutex_exit(&p->p_lock);
3212 mutex_exit(&stp->sd_lock);
3213 pgsignal(p->p_pgidp, SIGTTIN);
3214 mutex_enter(&stp->sd_lock);
3215 mutex_enter(&p->p_lock);
3216 } else { /* mode == JCWRITE or JCSETP */
3217 if ((mode == JCWRITE && !(stp->sd_flag & STRTOSTOP)) ||
3218 cantsend(p, t, SIGTTOU)) {
3219 mutex_exit(&p->p_lock);
3220 return (0);
3221 }
3222 if (p->p_detached) {
3223 mutex_exit(&p->p_lock);
3224 return (EIO);
3225 }
3226 mutex_exit(&p->p_lock);
3227 mutex_exit(&stp->sd_lock);
3228 pgsignal(p->p_pgidp, SIGTTOU);
3229 mutex_enter(&stp->sd_lock);
3230 mutex_enter(&p->p_lock);
3231 }
3232
3233 /*
3234 * We call cv_wait_sig_swap() to cause the appropriate
3235 * action for the jobcontrol signal to take place.
3236 * If the signal is being caught, we will take the
3237 * EINTR error return. Otherwise, the default action
3238 * of causing the process to stop will take place.
3239 * In this case, we rely on the periodic cv_broadcast() on
3240 * &lbolt_cv to wake us up to loop around and test again.
3241 * We can't get here if the signal is ignored or
3242 * if the current thread is blocking the signal.
3243 */
3244 mutex_exit(&stp->sd_lock);
3245 if (!cv_wait_sig_swap(&lbolt_cv, &p->p_lock)) {
3246 mutex_exit(&p->p_lock);
3247 mutex_enter(&stp->sd_lock);
3248 return (EINTR);
3249 }
3250 mutex_exit(&p->p_lock);
3251 mutex_enter(&stp->sd_lock);
3252 mutex_enter(&p->p_lock);
3253 }
3254 }
3255
3256 /*
3257 * Return size of message of block type (bp->b_datap->db_type)
3258 */
3259 size_t
3260 xmsgsize(mblk_t *bp)
3261 {
3262 unsigned char type;
3263 size_t count = 0;
3264
3265 type = bp->b_datap->db_type;
3266
3267 for (; bp; bp = bp->b_cont) {
3268 if (type != bp->b_datap->db_type)
3269 break;
3270 ASSERT(bp->b_wptr >= bp->b_rptr);
3271 count += bp->b_wptr - bp->b_rptr;
3272 }
3273 return (count);
3274 }
3275
3276 /*
3277 * Allocate a stream head.
3278 */
3279 struct stdata *
3280 shalloc(queue_t *qp)
3281 {
3282 stdata_t *stp;
3283
3284 stp = kmem_cache_alloc(stream_head_cache, KM_SLEEP);
3285
3286 stp->sd_wrq = _WR(qp);
3287 stp->sd_strtab = NULL;
3288 stp->sd_iocid = 0;
3289 stp->sd_mate = NULL;
3290 stp->sd_freezer = NULL;
3291 stp->sd_refcnt = 0;
3292 stp->sd_wakeq = 0;
3293 stp->sd_anchor = 0;
3294 stp->sd_struiowrq = NULL;
3295 stp->sd_struiordq = NULL;
3296 stp->sd_struiodnak = 0;
3297 stp->sd_struionak = NULL;
3298 stp->sd_t_audit_data = NULL;
3299 stp->sd_rput_opt = 0;
3300 stp->sd_wput_opt = 0;
3301 stp->sd_read_opt = 0;
3302 stp->sd_rprotofunc = strrput_proto;
3303 stp->sd_rmiscfunc = strrput_misc;
3304 stp->sd_rderrfunc = stp->sd_wrerrfunc = NULL;
3305 stp->sd_rputdatafunc = stp->sd_wputdatafunc = NULL;
3306 stp->sd_ciputctrl = NULL;
3307 stp->sd_nciputctrl = 0;
3308 stp->sd_qhead = NULL;
3309 stp->sd_qtail = NULL;
3310 stp->sd_servid = NULL;
3311 stp->sd_nqueues = 0;
3312 stp->sd_svcflags = 0;
3313 stp->sd_copyflag = 0;
3314
3315 return (stp);
3316 }
3317
3318 /*
3319 * Free a stream head.
3320 */
3321 void
3322 shfree(stdata_t *stp)
3323 {
3324 pid_node_t *pn;
3325
3326 ASSERT(MUTEX_NOT_HELD(&stp->sd_lock));
3327
3328 stp->sd_wrq = NULL;
3329
3330 mutex_enter(&stp->sd_qlock);
3331 while (stp->sd_svcflags & STRS_SCHEDULED) {
3332 STRSTAT(strwaits);
3333 cv_wait(&stp->sd_qcv, &stp->sd_qlock);
3334 }
3335 mutex_exit(&stp->sd_qlock);
3336
3337 if (stp->sd_ciputctrl != NULL) {
3338 ASSERT(stp->sd_nciputctrl == n_ciputctrl - 1);
3339 SUMCHECK_CIPUTCTRL_COUNTS(stp->sd_ciputctrl,
3340 stp->sd_nciputctrl, 0);
3341 ASSERT(ciputctrl_cache != NULL);
3342 kmem_cache_free(ciputctrl_cache, stp->sd_ciputctrl);
3343 stp->sd_ciputctrl = NULL;
3344 stp->sd_nciputctrl = 0;
3345 }
3346 ASSERT(stp->sd_qhead == NULL);
3347 ASSERT(stp->sd_qtail == NULL);
3348 ASSERT(stp->sd_nqueues == 0);
3349
3350 mutex_enter(&stp->sd_pid_list_lock);
3351 while ((pn = list_head(&stp->sd_pid_list)) != NULL) {
3352 list_remove(&stp->sd_pid_list, pn);
3353 kmem_free(pn, sizeof (*pn));
3354 }
3355 mutex_exit(&stp->sd_pid_list_lock);
3356
3357 kmem_cache_free(stream_head_cache, stp);
3358 }
3359
3360 void
3361 sh_insert_pid(struct stdata *stp, pid_t pid)
3362 {
3363 pid_node_t *pn;
3364
3365 mutex_enter(&stp->sd_pid_list_lock);
3366 for (pn = list_head(&stp->sd_pid_list);
3367 pn != NULL && pn->pn_pid != pid;
3368 pn = list_next(&stp->sd_pid_list, pn))
3369 ;
3370
3371 if (pn != NULL) {
3372 pn->pn_count++;
3373 } else {
3374 pn = kmem_zalloc(sizeof (*pn), KM_SLEEP);
3375 list_link_init(&pn->pn_ref_link);
3376 pn->pn_pid = pid;
3377 pn->pn_count = 1;
3378 list_insert_tail(&stp->sd_pid_list, pn);
3379 }
3380 mutex_exit(&stp->sd_pid_list_lock);
3381 }
3382
3383 void
3384 sh_remove_pid(struct stdata *stp, pid_t pid)
3385 {
3386 pid_node_t *pn;
3387
3388 mutex_enter(&stp->sd_pid_list_lock);
3389 for (pn = list_head(&stp->sd_pid_list);
3390 pn != NULL && pn->pn_pid != pid;
3391 pn = list_next(&stp->sd_pid_list, pn))
3392 ;
3393
3394 if (pn != NULL) {
3395 if (pn->pn_count > 1) {
3396 pn->pn_count--;
3397 } else {
3398 list_remove(&stp->sd_pid_list, pn);
3399 kmem_free(pn, sizeof (*pn));
3400 }
3401 }
3402 mutex_exit(&stp->sd_pid_list_lock);
3403 }
3404
3405 mblk_t *
3406 sh_get_pid_mblk(struct stdata *stp)
3407 {
3408 mblk_t *mblk;
3409 int sz, n = 0;
3410 pid_t *pids;
3411 pid_node_t *pn;
3412 conn_pid_info_t *cpi;
3413
3414 mutex_enter(&stp->sd_pid_list_lock);
3415
3416 n = list_numnodes(&stp->sd_pid_list);
3417 sz = sizeof (conn_pid_info_t);
3418 sz += (n > 1) ? ((n - 1) * sizeof (pid_t)) : 0;
3419 if ((mblk = allocb(sz, BPRI_HI)) == NULL) {
3420 mutex_exit(&stp->sd_pid_list_lock);
3421 return (NULL);
3422 }
3423 mblk->b_wptr += sz;
3424 cpi = (conn_pid_info_t *)mblk->b_datap->db_base;
3425 cpi->cpi_magic = CONN_PID_INFO_MGC;
3426 cpi->cpi_contents = CONN_PID_INFO_XTI;
3427 cpi->cpi_pids_cnt = n;
3428 cpi->cpi_tot_size = sz;
3429 cpi->cpi_pids[0] = 0;
3430
3431 if (cpi->cpi_pids_cnt > 0) {
3432 pids = cpi->cpi_pids;
3433 for (pn = list_head(&stp->sd_pid_list); pn != NULL;
3434 pids++, pn = list_next(&stp->sd_pid_list, pn))
3435 *pids = pn->pn_pid;
3436 }
3437 mutex_exit(&stp->sd_pid_list_lock);
3438 return (mblk);
3439 }
3440
3441 /*
3442 * Allocate a pair of queues and a syncq for the pair
3443 */
3444 queue_t *
3445 allocq(void)
3446 {
3447 queinfo_t *qip;
3448 queue_t *qp, *wqp;
3449 syncq_t *sq;
3450
3451 qip = kmem_cache_alloc(queue_cache, KM_SLEEP);
3452
3453 qp = &qip->qu_rqueue;
3454 wqp = &qip->qu_wqueue;
3455 sq = &qip->qu_syncq;
3456
3457 qp->q_last = NULL;
3458 qp->q_next = NULL;
3459 qp->q_ptr = NULL;
3460 qp->q_flag = QUSE | QREADR;
3461 qp->q_bandp = NULL;
3462 qp->q_stream = NULL;
3463 qp->q_syncq = sq;
3464 qp->q_nband = 0;
3465 qp->q_nfsrv = NULL;
3466 qp->q_draining = 0;
3467 qp->q_syncqmsgs = 0;
3468 qp->q_spri = 0;
3469 qp->q_qtstamp = 0;
3470 qp->q_sqtstamp = 0;
3471 qp->q_fp = NULL;
3472
3473 wqp->q_last = NULL;
3474 wqp->q_next = NULL;
3475 wqp->q_ptr = NULL;
3476 wqp->q_flag = QUSE;
3477 wqp->q_bandp = NULL;
3478 wqp->q_stream = NULL;
3479 wqp->q_syncq = sq;
3480 wqp->q_nband = 0;
3481 wqp->q_nfsrv = NULL;
3482 wqp->q_draining = 0;
3483 wqp->q_syncqmsgs = 0;
3484 wqp->q_qtstamp = 0;
3485 wqp->q_sqtstamp = 0;
3486 wqp->q_spri = 0;
3487
3488 sq->sq_count = 0;
3489 sq->sq_rmqcount = 0;
3490 sq->sq_flags = 0;
3491 sq->sq_type = 0;
3492 sq->sq_callbflags = 0;
3493 sq->sq_cancelid = 0;
3494 sq->sq_ciputctrl = NULL;
3495 sq->sq_nciputctrl = 0;
3496 sq->sq_needexcl = 0;
3497 sq->sq_svcflags = 0;
3498
3499 return (qp);
3500 }
3501
3502 /*
3503 * Free a pair of queues and the "attached" syncq.
3504 * Discard any messages left on the syncq(s), remove the syncq(s) from the
3505 * outer perimeter, and free the syncq(s) if they are not the "attached" syncq.
3506 */
3507 void
3508 freeq(queue_t *qp)
3509 {
3510 qband_t *qbp, *nqbp;
3511 syncq_t *sq, *outer;
3512 queue_t *wqp = _WR(qp);
3513
3514 ASSERT(qp->q_flag & QREADR);
3515
3516 /*
3517 * If a previously dispatched taskq job is scheduled to run
3518 * sync_service() or a service routine is scheduled for the
3519 * queues about to be freed, wait here until all service is
3520 * done on the queue and all associated queues and syncqs.
3521 */
3522 wait_svc(qp);
3523
3524 (void) flush_syncq(qp->q_syncq, qp);
3525 (void) flush_syncq(wqp->q_syncq, wqp);
3526 ASSERT(qp->q_syncqmsgs == 0 && wqp->q_syncqmsgs == 0);
3527
3528 /*
3529 * Flush the queues before q_next is set to NULL This is needed
3530 * in order to backenable any downstream queue before we go away.
3531 * Note: we are already removed from the stream so that the
3532 * backenabling will not cause any messages to be delivered to our
3533 * put procedures.
3534 */
3535 flushq(qp, FLUSHALL);
3536 flushq(wqp, FLUSHALL);
3537
3538 /* Tidy up - removeq only does a half-remove from stream */
3539 qp->q_next = wqp->q_next = NULL;
3540 ASSERT(!(qp->q_flag & QENAB));
3541 ASSERT(!(wqp->q_flag & QENAB));
3542
3543 outer = qp->q_syncq->sq_outer;
3544 if (outer != NULL) {
3545 outer_remove(outer, qp->q_syncq);
3546 if (wqp->q_syncq != qp->q_syncq)
3547 outer_remove(outer, wqp->q_syncq);
3548 }
3549 /*
3550 * Free any syncqs that are outside what allocq returned.
3551 */
3552 if (qp->q_syncq != SQ(qp) && !(qp->q_flag & QPERMOD))
3553 free_syncq(qp->q_syncq);
3554 if (qp->q_syncq != wqp->q_syncq && wqp->q_syncq != SQ(qp))
3555 free_syncq(wqp->q_syncq);
3556
3557 ASSERT((qp->q_sqflags & (Q_SQQUEUED | Q_SQDRAINING)) == 0);
3558 ASSERT((wqp->q_sqflags & (Q_SQQUEUED | Q_SQDRAINING)) == 0);
3559 ASSERT(MUTEX_NOT_HELD(QLOCK(qp)));
3560 ASSERT(MUTEX_NOT_HELD(QLOCK(wqp)));
3561 sq = SQ(qp);
3562 ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
3563 ASSERT(sq->sq_head == NULL && sq->sq_tail == NULL);
3564 ASSERT(sq->sq_outer == NULL);
3565 ASSERT(sq->sq_onext == NULL && sq->sq_oprev == NULL);
3566 ASSERT(sq->sq_callbpend == NULL);
3567 ASSERT(sq->sq_needexcl == 0);
3568
3569 if (sq->sq_ciputctrl != NULL) {
3570 ASSERT(sq->sq_nciputctrl == n_ciputctrl - 1);
3571 SUMCHECK_CIPUTCTRL_COUNTS(sq->sq_ciputctrl,
3572 sq->sq_nciputctrl, 0);
3573 ASSERT(ciputctrl_cache != NULL);
3574 kmem_cache_free(ciputctrl_cache, sq->sq_ciputctrl);
3575 sq->sq_ciputctrl = NULL;
3576 sq->sq_nciputctrl = 0;
3577 }
3578
3579 ASSERT(qp->q_first == NULL && wqp->q_first == NULL);
3580 ASSERT(qp->q_count == 0 && wqp->q_count == 0);
3581 ASSERT(qp->q_mblkcnt == 0 && wqp->q_mblkcnt == 0);
3582
3583 qp->q_flag &= ~QUSE;
3584 wqp->q_flag &= ~QUSE;
3585
3586 /* NOTE: Uncomment the assert below once bugid 1159635 is fixed. */
3587 /* ASSERT((qp->q_flag & QWANTW) == 0 && (wqp->q_flag & QWANTW) == 0); */
3588
3589 qbp = qp->q_bandp;
3590 while (qbp) {
3591 nqbp = qbp->qb_next;
3592 freeband(qbp);
3593 qbp = nqbp;
3594 }
3595 qbp = wqp->q_bandp;
3596 while (qbp) {
3597 nqbp = qbp->qb_next;
3598 freeband(qbp);
3599 qbp = nqbp;
3600 }
3601 kmem_cache_free(queue_cache, qp);
3602 }
3603
3604 /*
3605 * Allocate a qband structure.
3606 */
3607 qband_t *
3608 allocband(void)
3609 {
3610 qband_t *qbp;
3611
3612 qbp = kmem_cache_alloc(qband_cache, KM_NOSLEEP);
3613 if (qbp == NULL)
3614 return (NULL);
3615
3616 qbp->qb_next = NULL;
3617 qbp->qb_count = 0;
3618 qbp->qb_mblkcnt = 0;
3619 qbp->qb_first = NULL;
3620 qbp->qb_last = NULL;
3621 qbp->qb_flag = 0;
3622
3623 return (qbp);
3624 }
3625
3626 /*
3627 * Free a qband structure.
3628 */
3629 void
3630 freeband(qband_t *qbp)
3631 {
3632 kmem_cache_free(qband_cache, qbp);
3633 }
3634
3635 /*
3636 * Just like putnextctl(9F), except that allocb_wait() is used.
3637 *
3638 * Consolidation Private, and of course only callable from the stream head or
3639 * routines that may block.
3640 */
3641 int
3642 putnextctl_wait(queue_t *q, int type)
3643 {
3644 mblk_t *bp;
3645 int error;
3646
3647 if ((datamsg(type) && (type != M_DELAY)) ||
3648 (bp = allocb_wait(0, BPRI_HI, 0, &error)) == NULL)
3649 return (0);
3650
3651 bp->b_datap->db_type = (unsigned char)type;
3652 putnext(q, bp);
3653 return (1);
3654 }
3655
3656 /*
3657 * Run any possible bufcalls.
3658 */
3659 void
3660 runbufcalls(void)
3661 {
3662 strbufcall_t *bcp;
3663
3664 mutex_enter(&bcall_monitor);
3665 mutex_enter(&strbcall_lock);
3666
3667 if (strbcalls.bc_head) {
3668 size_t count;
3669 int nevent;
3670
3671 /*
3672 * count how many events are on the list
3673 * now so we can check to avoid looping
3674 * in low memory situations
3675 */
3676 nevent = 0;
3677 for (bcp = strbcalls.bc_head; bcp; bcp = bcp->bc_next)
3678 nevent++;
3679
3680 /*
3681 * get estimate of available memory from kmem_avail().
3682 * awake all bufcall functions waiting for
3683 * memory whose request could be satisfied
3684 * by 'count' memory and let 'em fight for it.
3685 */
3686 count = kmem_avail();
3687 while ((bcp = strbcalls.bc_head) != NULL && nevent) {
3688 STRSTAT(bufcalls);
3689 --nevent;
3690 if (bcp->bc_size <= count) {
3691 bcp->bc_executor = curthread;
3692 mutex_exit(&strbcall_lock);
3693 (*bcp->bc_func)(bcp->bc_arg);
3694 mutex_enter(&strbcall_lock);
3695 bcp->bc_executor = NULL;
3696 cv_broadcast(&bcall_cv);
3697 strbcalls.bc_head = bcp->bc_next;
3698 kmem_free(bcp, sizeof (strbufcall_t));
3699 } else {
3700 /*
3701 * too big, try again later - note
3702 * that nevent was decremented above
3703 * so we won't retry this one on this
3704 * iteration of the loop
3705 */
3706 if (bcp->bc_next != NULL) {
3707 strbcalls.bc_head = bcp->bc_next;
3708 bcp->bc_next = NULL;
3709 strbcalls.bc_tail->bc_next = bcp;
3710 strbcalls.bc_tail = bcp;
3711 }
3712 }
3713 }
3714 if (strbcalls.bc_head == NULL)
3715 strbcalls.bc_tail = NULL;
3716 }
3717
3718 mutex_exit(&strbcall_lock);
3719 mutex_exit(&bcall_monitor);
3720 }
3721
3722
3723 /*
3724 * Actually run queue's service routine.
3725 */
3726 static void
3727 runservice(queue_t *q)
3728 {
3729 qband_t *qbp;
3730
3731 ASSERT(q->q_qinfo->qi_srvp);
3732 again:
3733 entersq(q->q_syncq, SQ_SVC);
3734 TRACE_1(TR_FAC_STREAMS_FR, TR_QRUNSERVICE_START,
3735 "runservice starts:%p", q);
3736
3737 if (!(q->q_flag & QWCLOSE))
3738 (*q->q_qinfo->qi_srvp)(q);
3739
3740 TRACE_1(TR_FAC_STREAMS_FR, TR_QRUNSERVICE_END,
3741 "runservice ends:(%p)", q);
3742
3743 leavesq(q->q_syncq, SQ_SVC);
3744
3745 mutex_enter(QLOCK(q));
3746 if (q->q_flag & QENAB) {
3747 q->q_flag &= ~QENAB;
3748 mutex_exit(QLOCK(q));
3749 goto again;
3750 }
3751 q->q_flag &= ~QINSERVICE;
3752 q->q_flag &= ~QBACK;
3753 for (qbp = q->q_bandp; qbp; qbp = qbp->qb_next)
3754 qbp->qb_flag &= ~QB_BACK;
3755 /*
3756 * Wakeup thread waiting for the service procedure
3757 * to be run (strclose and qdetach).
3758 */
3759 cv_broadcast(&q->q_wait);
3760
3761 mutex_exit(QLOCK(q));
3762 }
3763
3764 /*
3765 * Background processing of bufcalls.
3766 */
3767 void
3768 streams_bufcall_service(void)
3769 {
3770 callb_cpr_t cprinfo;
3771
3772 CALLB_CPR_INIT(&cprinfo, &strbcall_lock, callb_generic_cpr,
3773 "streams_bufcall_service");
3774
3775 mutex_enter(&strbcall_lock);
3776
3777 for (;;) {
3778 if (strbcalls.bc_head != NULL && kmem_avail() > 0) {
3779 mutex_exit(&strbcall_lock);
3780 runbufcalls();
3781 mutex_enter(&strbcall_lock);
3782 }
3783 if (strbcalls.bc_head != NULL) {
3784 STRSTAT(bcwaits);
3785 /* Wait for memory to become available */
3786 CALLB_CPR_SAFE_BEGIN(&cprinfo);
3787 (void) cv_reltimedwait(&memavail_cv, &strbcall_lock,
3788 SEC_TO_TICK(60), TR_CLOCK_TICK);
3789 CALLB_CPR_SAFE_END(&cprinfo, &strbcall_lock);
3790 }
3791
3792 /* Wait for new work to arrive */
3793 if (strbcalls.bc_head == NULL) {
3794 CALLB_CPR_SAFE_BEGIN(&cprinfo);
3795 cv_wait(&strbcall_cv, &strbcall_lock);
3796 CALLB_CPR_SAFE_END(&cprinfo, &strbcall_lock);
3797 }
3798 }
3799 }
3800
3801 /*
3802 * Background processing of streams background tasks which failed
3803 * taskq_dispatch.
3804 */
3805 static void
3806 streams_qbkgrnd_service(void)
3807 {
3808 callb_cpr_t cprinfo;
3809 queue_t *q;
3810
3811 CALLB_CPR_INIT(&cprinfo, &service_queue, callb_generic_cpr,
3812 "streams_bkgrnd_service");
3813
3814 mutex_enter(&service_queue);
3815
3816 for (;;) {
3817 /*
3818 * Wait for work to arrive.
3819 */
3820 while ((freebs_list == NULL) && (qhead == NULL)) {
3821 CALLB_CPR_SAFE_BEGIN(&cprinfo);
3822 cv_wait(&services_to_run, &service_queue);
3823 CALLB_CPR_SAFE_END(&cprinfo, &service_queue);
3824 }
3825 /*
3826 * Handle all pending freebs requests to free memory.
3827 */
3828 while (freebs_list != NULL) {
3829 mblk_t *mp = freebs_list;
3830 freebs_list = mp->b_next;
3831 mutex_exit(&service_queue);
3832 mblk_free(mp);
3833 mutex_enter(&service_queue);
3834 }
3835 /*
3836 * Run pending queues.
3837 */
3838 while (qhead != NULL) {
3839 DQ(q, qhead, qtail, q_link);
3840 ASSERT(q != NULL);
3841 mutex_exit(&service_queue);
3842 queue_service(q);
3843 mutex_enter(&service_queue);
3844 }
3845 ASSERT(qhead == NULL && qtail == NULL);
3846 }
3847 }
3848
3849 /*
3850 * Background processing of streams background tasks which failed
3851 * taskq_dispatch.
3852 */
3853 static void
3854 streams_sqbkgrnd_service(void)
3855 {
3856 callb_cpr_t cprinfo;
3857 syncq_t *sq;
3858
3859 CALLB_CPR_INIT(&cprinfo, &service_queue, callb_generic_cpr,
3860 "streams_sqbkgrnd_service");
3861
3862 mutex_enter(&service_queue);
3863
3864 for (;;) {
3865 /*
3866 * Wait for work to arrive.
3867 */
3868 while (sqhead == NULL) {
3869 CALLB_CPR_SAFE_BEGIN(&cprinfo);
3870 cv_wait(&syncqs_to_run, &service_queue);
3871 CALLB_CPR_SAFE_END(&cprinfo, &service_queue);
3872 }
3873
3874 /*
3875 * Run pending syncqs.
3876 */
3877 while (sqhead != NULL) {
3878 DQ(sq, sqhead, sqtail, sq_next);
3879 ASSERT(sq != NULL);
3880 ASSERT(sq->sq_svcflags & SQ_BGTHREAD);
3881 mutex_exit(&service_queue);
3882 syncq_service(sq);
3883 mutex_enter(&service_queue);
3884 }
3885 }
3886 }
3887
3888 /*
3889 * Disable the syncq and wait for background syncq processing to complete.
3890 * If the syncq is placed on the sqhead/sqtail queue, try to remove it from the
3891 * list.
3892 */
3893 void
3894 wait_sq_svc(syncq_t *sq)
3895 {
3896 mutex_enter(SQLOCK(sq));
3897 sq->sq_svcflags |= SQ_DISABLED;
3898 if (sq->sq_svcflags & SQ_BGTHREAD) {
3899 syncq_t *sq_chase;
3900 syncq_t *sq_curr;
3901 int removed;
3902
3903 ASSERT(sq->sq_servcount == 1);
3904 mutex_enter(&service_queue);
3905 RMQ(sq, sqhead, sqtail, sq_next, sq_chase, sq_curr, removed);
3906 mutex_exit(&service_queue);
3907 if (removed) {
3908 sq->sq_svcflags &= ~SQ_BGTHREAD;
3909 sq->sq_servcount = 0;
3910 STRSTAT(sqremoved);
3911 goto done;
3912 }
3913 }
3914 while (sq->sq_servcount != 0) {
3915 sq->sq_flags |= SQ_WANTWAKEUP;
3916 cv_wait(&sq->sq_wait, SQLOCK(sq));
3917 }
3918 done:
3919 mutex_exit(SQLOCK(sq));
3920 }
3921
3922 /*
3923 * Put a syncq on the list of syncq's to be serviced by the sqthread.
3924 * Add the argument to the end of the sqhead list and set the flag
3925 * indicating this syncq has been enabled. If it has already been
3926 * enabled, don't do anything.
3927 * This routine assumes that SQLOCK is held.
3928 * NOTE that the lock order is to have the SQLOCK first,
3929 * so if the service_syncq lock is held, we need to release it
3930 * before acquiring the SQLOCK (mostly relevant for the background
3931 * thread, and this seems to be common among the STREAMS global locks).
3932 * Note that the sq_svcflags are protected by the SQLOCK.
3933 */
3934 void
3935 sqenable(syncq_t *sq)
3936 {
3937 /*
3938 * This is probably not important except for where I believe it
3939 * is being called. At that point, it should be held (and it
3940 * is a pain to release it just for this routine, so don't do
3941 * it).
3942 */
3943 ASSERT(MUTEX_HELD(SQLOCK(sq)));
3944
3945 IMPLY(sq->sq_servcount == 0, sq->sq_next == NULL);
3946 IMPLY(sq->sq_next != NULL, sq->sq_svcflags & SQ_BGTHREAD);
3947
3948 /*
3949 * Do not put on list if background thread is scheduled or
3950 * syncq is disabled.
3951 */
3952 if (sq->sq_svcflags & (SQ_DISABLED | SQ_BGTHREAD))
3953 return;
3954
3955 /*
3956 * Check whether we should enable sq at all.
3957 * Non PERMOD syncqs may be drained by at most one thread.
3958 * PERMOD syncqs may be drained by several threads but we limit the
3959 * total amount to the lesser of
3960 * Number of queues on the squeue and
3961 * Number of CPUs.
3962 */
3963 if (sq->sq_servcount != 0) {
3964 if (((sq->sq_type & SQ_PERMOD) == 0) ||
3965 (sq->sq_servcount >= MIN(sq->sq_nqueues, ncpus_online))) {
3966 STRSTAT(sqtoomany);
3967 return;
3968 }
3969 }
3970
3971 sq->sq_tstamp = ddi_get_lbolt();
3972 STRSTAT(sqenables);
3973
3974 /* Attempt a taskq dispatch */
3975 sq->sq_servid = (void *)taskq_dispatch(streams_taskq,
3976 (task_func_t *)syncq_service, sq, TQ_NOSLEEP | TQ_NOQUEUE);
3977 if (sq->sq_servid != NULL) {
3978 sq->sq_servcount++;
3979 return;
3980 }
3981
3982 /*
3983 * This taskq dispatch failed, but a previous one may have succeeded.
3984 * Don't try to schedule on the background thread whilst there is
3985 * outstanding taskq processing.
3986 */
3987 if (sq->sq_servcount != 0)
3988 return;
3989
3990 /*
3991 * System is low on resources and can't perform a non-sleeping
3992 * dispatch. Schedule the syncq for a background thread and mark the
3993 * syncq to avoid any further taskq dispatch attempts.
3994 */
3995 mutex_enter(&service_queue);
3996 STRSTAT(taskqfails);
3997 ENQUEUE(sq, sqhead, sqtail, sq_next);
3998 sq->sq_svcflags |= SQ_BGTHREAD;
3999 sq->sq_servcount = 1;
4000 cv_signal(&syncqs_to_run);
4001 mutex_exit(&service_queue);
4002 }
4003
4004 /*
4005 * Note: fifo_close() depends on the mblk_t on the queue being freed
4006 * asynchronously. The asynchronous freeing of messages breaks the
4007 * recursive call chain of fifo_close() while there are I_SENDFD type of
4008 * messages referring to other file pointers on the queue. Then when
4009 * closing pipes it can avoid stack overflow in case of daisy-chained
4010 * pipes, and also avoid deadlock in case of fifonode_t pairs (which
4011 * share the same fifolock_t).
4012 *
4013 * No need to kpreempt_disable to access cpu_seqid. If we migrate and
4014 * the esb queue does not match the new CPU, that is OK.
4015 */
4016 void
4017 freebs_enqueue(mblk_t *mp, dblk_t *dbp)
4018 {
4019 int qindex = CPU->cpu_seqid >> esbq_log2_cpus_per_q;
4020 esb_queue_t *eqp;
4021
4022 ASSERT(dbp->db_mblk == mp);
4023 ASSERT(qindex < esbq_nelem);
4024
4025 eqp = system_esbq_array;
4026 if (eqp != NULL) {
4027 eqp += qindex;
4028 } else {
4029 mutex_enter(&esbq_lock);
4030 if (kmem_ready && system_esbq_array == NULL)
4031 system_esbq_array = (esb_queue_t *)kmem_zalloc(
4032 esbq_nelem * sizeof (esb_queue_t), KM_NOSLEEP);
4033 mutex_exit(&esbq_lock);
4034 eqp = system_esbq_array;
4035 if (eqp != NULL)
4036 eqp += qindex;
4037 else
4038 eqp = &system_esbq;
4039 }
4040
4041 /*
4042 * Check data sanity. The dblock should have non-empty free function.
4043 * It is better to panic here then later when the dblock is freed
4044 * asynchronously when the context is lost.
4045 */
4046 if (dbp->db_frtnp->free_func == NULL) {
4047 panic("freebs_enqueue: dblock %p has a NULL free callback",
4048 (void *)dbp);
4049 }
4050
4051 mutex_enter(&eqp->eq_lock);
4052 /* queue the new mblk on the esballoc queue */
4053 if (eqp->eq_head == NULL) {
4054 eqp->eq_head = eqp->eq_tail = mp;
4055 } else {
4056 eqp->eq_tail->b_next = mp;
4057 eqp->eq_tail = mp;
4058 }
4059 eqp->eq_len++;
4060
4061 /* If we're the first thread to reach the threshold, process */
4062 if (eqp->eq_len >= esbq_max_qlen &&
4063 !(eqp->eq_flags & ESBQ_PROCESSING))
4064 esballoc_process_queue(eqp);
4065
4066 esballoc_set_timer(eqp, esbq_timeout);
4067 mutex_exit(&eqp->eq_lock);
4068 }
4069
4070 static void
4071 esballoc_process_queue(esb_queue_t *eqp)
4072 {
4073 mblk_t *mp;
4074
4075 ASSERT(MUTEX_HELD(&eqp->eq_lock));
4076
4077 eqp->eq_flags |= ESBQ_PROCESSING;
4078
4079 do {
4080 /*
4081 * Detach the message chain for processing.
4082 */
4083 mp = eqp->eq_head;
4084 eqp->eq_tail->b_next = NULL;
4085 eqp->eq_head = eqp->eq_tail = NULL;
4086 eqp->eq_len = 0;
4087 mutex_exit(&eqp->eq_lock);
4088
4089 /*
4090 * Process the message chain.
4091 */
4092 esballoc_enqueue_mblk(mp);
4093 mutex_enter(&eqp->eq_lock);
4094 } while ((eqp->eq_len >= esbq_max_qlen) && (eqp->eq_len > 0));
4095
4096 eqp->eq_flags &= ~ESBQ_PROCESSING;
4097 }
4098
4099 /*
4100 * taskq callback routine to free esballoced mblk's
4101 */
4102 static void
4103 esballoc_mblk_free(mblk_t *mp)
4104 {
4105 mblk_t *nextmp;
4106
4107 for (; mp != NULL; mp = nextmp) {
4108 nextmp = mp->b_next;
4109 mp->b_next = NULL;
4110 mblk_free(mp);
4111 }
4112 }
4113
4114 static void
4115 esballoc_enqueue_mblk(mblk_t *mp)
4116 {
4117
4118 if (taskq_dispatch(system_taskq, (task_func_t *)esballoc_mblk_free, mp,
4119 TQ_NOSLEEP) == NULL) {
4120 mblk_t *first_mp = mp;
4121 /*
4122 * System is low on resources and can't perform a non-sleeping
4123 * dispatch. Schedule for a background thread.
4124 */
4125 mutex_enter(&service_queue);
4126 STRSTAT(taskqfails);
4127
4128 while (mp->b_next != NULL)
4129 mp = mp->b_next;
4130
4131 mp->b_next = freebs_list;
4132 freebs_list = first_mp;
4133 cv_signal(&services_to_run);
4134 mutex_exit(&service_queue);
4135 }
4136 }
4137
4138 static void
4139 esballoc_timer(void *arg)
4140 {
4141 esb_queue_t *eqp = arg;
4142
4143 mutex_enter(&eqp->eq_lock);
4144 eqp->eq_flags &= ~ESBQ_TIMER;
4145
4146 if (!(eqp->eq_flags & ESBQ_PROCESSING) &&
4147 eqp->eq_len > 0)
4148 esballoc_process_queue(eqp);
4149
4150 esballoc_set_timer(eqp, esbq_timeout);
4151 mutex_exit(&eqp->eq_lock);
4152 }
4153
4154 static void
4155 esballoc_set_timer(esb_queue_t *eqp, clock_t eq_timeout)
4156 {
4157 ASSERT(MUTEX_HELD(&eqp->eq_lock));
4158
4159 if (eqp->eq_len > 0 && !(eqp->eq_flags & ESBQ_TIMER)) {
4160 (void) timeout(esballoc_timer, eqp, eq_timeout);
4161 eqp->eq_flags |= ESBQ_TIMER;
4162 }
4163 }
4164
4165 /*
4166 * Setup esbq array length based upon NCPU scaled by CPUs per
4167 * queue. Use static system_esbq until kmem_ready and we can
4168 * create an array in freebs_enqueue().
4169 */
4170 void
4171 esballoc_queue_init(void)
4172 {
4173 esbq_log2_cpus_per_q = highbit(esbq_cpus_per_q - 1);
4174 esbq_cpus_per_q = 1 << esbq_log2_cpus_per_q;
4175 esbq_nelem = howmany(NCPU, esbq_cpus_per_q);
4176 system_esbq.eq_len = 0;
4177 system_esbq.eq_head = system_esbq.eq_tail = NULL;
4178 system_esbq.eq_flags = 0;
4179 }
4180
4181 /*
4182 * Set the QBACK or QB_BACK flag in the given queue for
4183 * the given priority band.
4184 */
4185 void
4186 setqback(queue_t *q, unsigned char pri)
4187 {
4188 int i;
4189 qband_t *qbp;
4190 qband_t **qbpp;
4191
4192 ASSERT(MUTEX_HELD(QLOCK(q)));
4193 if (pri != 0) {
4194 if (pri > q->q_nband) {
4195 qbpp = &q->q_bandp;
4196 while (*qbpp)
4197 qbpp = &(*qbpp)->qb_next;
4198 while (pri > q->q_nband) {
4199 if ((*qbpp = allocband()) == NULL) {
4200 cmn_err(CE_WARN,
4201 "setqback: can't allocate qband\n");
4202 return;
4203 }
4204 (*qbpp)->qb_hiwat = q->q_hiwat;
4205 (*qbpp)->qb_lowat = q->q_lowat;
4206 q->q_nband++;
4207 qbpp = &(*qbpp)->qb_next;
4208 }
4209 }
4210 qbp = q->q_bandp;
4211 i = pri;
4212 while (--i)
4213 qbp = qbp->qb_next;
4214 qbp->qb_flag |= QB_BACK;
4215 } else {
4216 q->q_flag |= QBACK;
4217 }
4218 }
4219
4220 int
4221 strcopyin(void *from, void *to, size_t len, int copyflag)
4222 {
4223 if (copyflag & U_TO_K) {
4224 ASSERT((copyflag & K_TO_K) == 0);
4225 if (copyin(from, to, len))
4226 return (EFAULT);
4227 } else {
4228 ASSERT(copyflag & K_TO_K);
4229 bcopy(from, to, len);
4230 }
4231 return (0);
4232 }
4233
4234 int
4235 strcopyout(void *from, void *to, size_t len, int copyflag)
4236 {
4237 if (copyflag & U_TO_K) {
4238 if (copyout(from, to, len))
4239 return (EFAULT);
4240 } else {
4241 ASSERT(copyflag & K_TO_K);
4242 bcopy(from, to, len);
4243 }
4244 return (0);
4245 }
4246
4247 /*
4248 * strsignal_nolock() posts a signal to the process(es) at the stream head.
4249 * It assumes that the stream head lock is already held, whereas strsignal()
4250 * acquires the lock first. This routine was created because a few callers
4251 * release the stream head lock before calling only to re-acquire it after
4252 * it returns.
4253 */
4254 void
4255 strsignal_nolock(stdata_t *stp, int sig, uchar_t band)
4256 {
4257 ASSERT(MUTEX_HELD(&stp->sd_lock));
4258 switch (sig) {
4259 case SIGPOLL:
4260 if (stp->sd_sigflags & S_MSG)
4261 strsendsig(stp->sd_siglist, S_MSG, band, 0);
4262 break;
4263 default:
4264 if (stp->sd_pgidp)
4265 pgsignal(stp->sd_pgidp, sig);
4266 break;
4267 }
4268 }
4269
4270 void
4271 strsignal(stdata_t *stp, int sig, int32_t band)
4272 {
4273 TRACE_3(TR_FAC_STREAMS_FR, TR_SENDSIG,
4274 "strsignal:%p, %X, %X", stp, sig, band);
4275
4276 mutex_enter(&stp->sd_lock);
4277 switch (sig) {
4278 case SIGPOLL:
4279 if (stp->sd_sigflags & S_MSG)
4280 strsendsig(stp->sd_siglist, S_MSG, (uchar_t)band, 0);
4281 break;
4282
4283 default:
4284 if (stp->sd_pgidp) {
4285 pgsignal(stp->sd_pgidp, sig);
4286 }
4287 break;
4288 }
4289 mutex_exit(&stp->sd_lock);
4290 }
4291
4292 void
4293 strhup(stdata_t *stp)
4294 {
4295 ASSERT(mutex_owned(&stp->sd_lock));
4296 pollwakeup(&stp->sd_pollist, POLLHUP);
4297 if (stp->sd_sigflags & S_HANGUP)
4298 strsendsig(stp->sd_siglist, S_HANGUP, 0, 0);
4299 }
4300
4301 /*
4302 * Backenable the first queue upstream from `q' with a service procedure.
4303 */
4304 void
4305 backenable(queue_t *q, uchar_t pri)
4306 {
4307 queue_t *nq;
4308
4309 /*
4310 * Our presence might not prevent other modules in our own
4311 * stream from popping/pushing since the caller of getq might not
4312 * have a claim on the queue (some drivers do a getq on somebody
4313 * else's queue - they know that the queue itself is not going away
4314 * but the framework has to guarantee q_next in that stream).
4315 */
4316 claimstr(q);
4317
4318 /* Find nearest back queue with service proc */
4319 for (nq = backq(q); nq && !nq->q_qinfo->qi_srvp; nq = backq(nq)) {
4320 ASSERT(STRMATED(q->q_stream) || STREAM(q) == STREAM(nq));
4321 }
4322
4323 if (nq) {
4324 kthread_t *freezer;
4325 /*
4326 * backenable can be called either with no locks held
4327 * or with the stream frozen (the latter occurs when a module
4328 * calls rmvq with the stream frozen). If the stream is frozen
4329 * by the caller the caller will hold all qlocks in the stream.
4330 * Note that a frozen stream doesn't freeze a mated stream,
4331 * so we explicitly check for that.
4332 */
4333 freezer = STREAM(q)->sd_freezer;
4334 if (freezer != curthread || STREAM(q) != STREAM(nq)) {
4335 mutex_enter(QLOCK(nq));
4336 }
4337 #ifdef DEBUG
4338 else {
4339 ASSERT(frozenstr(q));
4340 ASSERT(MUTEX_HELD(QLOCK(q)));
4341 ASSERT(MUTEX_HELD(QLOCK(nq)));
4342 }
4343 #endif
4344 setqback(nq, pri);
4345 qenable_locked(nq);
4346 if (freezer != curthread || STREAM(q) != STREAM(nq))
4347 mutex_exit(QLOCK(nq));
4348 }
4349 releasestr(q);
4350 }
4351
4352 /*
4353 * Return the appropriate errno when one of flags_to_check is set
4354 * in sd_flags. Uses the exported error routines if they are set.
4355 * Will return 0 if non error is set (or if the exported error routines
4356 * do not return an error).
4357 *
4358 * If there is both a read and write error to check, we prefer the read error.
4359 * Also, give preference to recorded errno's over the error functions.
4360 * The flags that are handled are:
4361 * STPLEX return EINVAL
4362 * STRDERR return sd_rerror (and clear if STRDERRNONPERSIST)
4363 * STWRERR return sd_werror (and clear if STWRERRNONPERSIST)
4364 * STRHUP return sd_werror
4365 *
4366 * If the caller indicates that the operation is a peek, a nonpersistent error
4367 * is not cleared.
4368 */
4369 int
4370 strgeterr(stdata_t *stp, int32_t flags_to_check, int ispeek)
4371 {
4372 int32_t sd_flag = stp->sd_flag & flags_to_check;
4373 int error = 0;
4374
4375 ASSERT(MUTEX_HELD(&stp->sd_lock));
4376 ASSERT((flags_to_check & ~(STRDERR|STWRERR|STRHUP|STPLEX)) == 0);
4377 if (sd_flag & STPLEX)
4378 error = EINVAL;
4379 else if (sd_flag & STRDERR) {
4380 error = stp->sd_rerror;
4381 if ((stp->sd_flag & STRDERRNONPERSIST) && !ispeek) {
4382 /*
4383 * Read errors are non-persistent i.e. discarded once
4384 * returned to a non-peeking caller,
4385 */
4386 stp->sd_rerror = 0;
4387 stp->sd_flag &= ~STRDERR;
4388 }
4389 if (error == 0 && stp->sd_rderrfunc != NULL) {
4390 int clearerr = 0;
4391
4392 error = (*stp->sd_rderrfunc)(stp->sd_vnode, ispeek,
4393 &clearerr);
4394 if (clearerr) {
4395 stp->sd_flag &= ~STRDERR;
4396 stp->sd_rderrfunc = NULL;
4397 }
4398 }
4399 } else if (sd_flag & STWRERR) {
4400 error = stp->sd_werror;
4401 if ((stp->sd_flag & STWRERRNONPERSIST) && !ispeek) {
4402 /*
4403 * Write errors are non-persistent i.e. discarded once
4404 * returned to a non-peeking caller,
4405 */
4406 stp->sd_werror = 0;
4407 stp->sd_flag &= ~STWRERR;
4408 }
4409 if (error == 0 && stp->sd_wrerrfunc != NULL) {
4410 int clearerr = 0;
4411
4412 error = (*stp->sd_wrerrfunc)(stp->sd_vnode, ispeek,
4413 &clearerr);
4414 if (clearerr) {
4415 stp->sd_flag &= ~STWRERR;
4416 stp->sd_wrerrfunc = NULL;
4417 }
4418 }
4419 } else if (sd_flag & STRHUP) {
4420 /* sd_werror set when STRHUP */
4421 error = stp->sd_werror;
4422 }
4423 return (error);
4424 }
4425
4426
4427 /*
4428 * Single-thread open/close/push/pop
4429 * for twisted streams also
4430 */
4431 int
4432 strstartplumb(stdata_t *stp, int flag, int cmd)
4433 {
4434 int waited = 1;
4435 int error = 0;
4436
4437 if (STRMATED(stp)) {
4438 struct stdata *stmatep = stp->sd_mate;
4439
4440 STRLOCKMATES(stp);
4441 while (waited) {
4442 waited = 0;
4443 while (stmatep->sd_flag & (STWOPEN|STRCLOSE|STRPLUMB)) {
4444 if ((cmd == I_POP) &&
4445 (flag & (FNDELAY|FNONBLOCK))) {
4446 STRUNLOCKMATES(stp);
4447 return (EAGAIN);
4448 }
4449 waited = 1;
4450 mutex_exit(&stp->sd_lock);
4451 if (!cv_wait_sig(&stmatep->sd_monitor,
4452 &stmatep->sd_lock)) {
4453 mutex_exit(&stmatep->sd_lock);
4454 return (EINTR);
4455 }
4456 mutex_exit(&stmatep->sd_lock);
4457 STRLOCKMATES(stp);
4458 }
4459 while (stp->sd_flag & (STWOPEN|STRCLOSE|STRPLUMB)) {
4460 if ((cmd == I_POP) &&
4461 (flag & (FNDELAY|FNONBLOCK))) {
4462 STRUNLOCKMATES(stp);
4463 return (EAGAIN);
4464 }
4465 waited = 1;
4466 mutex_exit(&stmatep->sd_lock);
4467 if (!cv_wait_sig(&stp->sd_monitor,
4468 &stp->sd_lock)) {
4469 mutex_exit(&stp->sd_lock);
4470 return (EINTR);
4471 }
4472 mutex_exit(&stp->sd_lock);
4473 STRLOCKMATES(stp);
4474 }
4475 if (stp->sd_flag & (STRDERR|STWRERR|STRHUP|STPLEX)) {
4476 error = strgeterr(stp,
4477 STRDERR|STWRERR|STRHUP|STPLEX, 0);
4478 if (error != 0) {
4479 STRUNLOCKMATES(stp);
4480 return (error);
4481 }
4482 }
4483 }
4484 stp->sd_flag |= STRPLUMB;
4485 STRUNLOCKMATES(stp);
4486 } else {
4487 mutex_enter(&stp->sd_lock);
4488 while (stp->sd_flag & (STWOPEN|STRCLOSE|STRPLUMB)) {
4489 if (((cmd == I_POP) || (cmd == _I_REMOVE)) &&
4490 (flag & (FNDELAY|FNONBLOCK))) {
4491 mutex_exit(&stp->sd_lock);
4492 return (EAGAIN);
4493 }
4494 if (!cv_wait_sig(&stp->sd_monitor, &stp->sd_lock)) {
4495 mutex_exit(&stp->sd_lock);
4496 return (EINTR);
4497 }
4498 if (stp->sd_flag & (STRDERR|STWRERR|STRHUP|STPLEX)) {
4499 error = strgeterr(stp,
4500 STRDERR|STWRERR|STRHUP|STPLEX, 0);
4501 if (error != 0) {
4502 mutex_exit(&stp->sd_lock);
4503 return (error);
4504 }
4505 }
4506 }
4507 stp->sd_flag |= STRPLUMB;
4508 mutex_exit(&stp->sd_lock);
4509 }
4510 return (0);
4511 }
4512
4513 /*
4514 * Complete the plumbing operation associated with stream `stp'.
4515 */
4516 void
4517 strendplumb(stdata_t *stp)
4518 {
4519 ASSERT(MUTEX_HELD(&stp->sd_lock));
4520 ASSERT(stp->sd_flag & STRPLUMB);
4521 stp->sd_flag &= ~STRPLUMB;
4522 cv_broadcast(&stp->sd_monitor);
4523 }
4524
4525 /*
4526 * This describes how the STREAMS framework handles synchronization
4527 * during open/push and close/pop.
4528 * The key interfaces for open and close are qprocson and qprocsoff,
4529 * respectively. While the close case in general is harder both open
4530 * have close have significant similarities.
4531 *
4532 * During close the STREAMS framework has to both ensure that there
4533 * are no stale references to the queue pair (and syncq) that
4534 * are being closed and also provide the guarantees that are documented
4535 * in qprocsoff(9F).
4536 * If there are stale references to the queue that is closing it can
4537 * result in kernel memory corruption or kernel panics.
4538 *
4539 * Note that is it up to the module/driver to ensure that it itself
4540 * does not have any stale references to the closing queues once its close
4541 * routine returns. This includes:
4542 * - Cancelling any timeout/bufcall/qtimeout/qbufcall callback routines
4543 * associated with the queues. For timeout and bufcall callbacks the
4544 * module/driver also has to ensure (or wait for) any callbacks that
4545 * are in progress.
4546 * - If the module/driver is using esballoc it has to ensure that any
4547 * esballoc free functions do not refer to a queue that has closed.
4548 * (Note that in general the close routine can not wait for the esballoc'ed
4549 * messages to be freed since that can cause a deadlock.)
4550 * - Cancelling any interrupts that refer to the closing queues and
4551 * also ensuring that there are no interrupts in progress that will
4552 * refer to the closing queues once the close routine returns.
4553 * - For multiplexors removing any driver global state that refers to
4554 * the closing queue and also ensuring that there are no threads in
4555 * the multiplexor that has picked up a queue pointer but not yet
4556 * finished using it.
4557 *
4558 * In addition, a driver/module can only reference the q_next pointer
4559 * in its open, close, put, or service procedures or in a
4560 * qtimeout/qbufcall callback procedure executing "on" the correct
4561 * stream. Thus it can not reference the q_next pointer in an interrupt
4562 * routine or a timeout, bufcall or esballoc callback routine. Likewise
4563 * it can not reference q_next of a different queue e.g. in a mux that
4564 * passes messages from one queues put/service procedure to another queue.
4565 * In all the cases when the driver/module can not access the q_next
4566 * field it must use the *next* versions e.g. canputnext instead of
4567 * canput(q->q_next) and putnextctl instead of putctl(q->q_next, ...).
4568 *
4569 *
4570 * Assuming that the driver/module conforms to the above constraints
4571 * the STREAMS framework has to avoid stale references to q_next for all
4572 * the framework internal cases which include (but are not limited to):
4573 * - Threads in canput/canputnext/backenable and elsewhere that are
4574 * walking q_next.
4575 * - Messages on a syncq that have a reference to the queue through b_queue.
4576 * - Messages on an outer perimeter (syncq) that have a reference to the
4577 * queue through b_queue.
4578 * - Threads that use q_nfsrv (e.g. canput) to find a queue.
4579 * Note that only canput and bcanput use q_nfsrv without any locking.
4580 *
4581 * The STREAMS framework providing the qprocsoff(9F) guarantees means that
4582 * after qprocsoff returns, the framework has to ensure that no threads can
4583 * enter the put or service routines for the closing read or write-side queue.
4584 * In addition to preventing "direct" entry into the put procedures
4585 * the framework also has to prevent messages being drained from
4586 * the syncq or the outer perimeter.
4587 * XXX Note that currently qdetach does relies on D_MTOCEXCL as the only
4588 * mechanism to prevent qwriter(PERIM_OUTER) from running after
4589 * qprocsoff has returned.
4590 * Note that if a module/driver uses put(9F) on one of its own queues
4591 * it is up to the module/driver to ensure that the put() doesn't
4592 * get called when the queue is closing.
4593 *
4594 *
4595 * The framework aspects of the above "contract" is implemented by
4596 * qprocsoff, removeq, and strlock:
4597 * - qprocsoff (disable_svc) sets QWCLOSE to prevent runservice from
4598 * entering the service procedures.
4599 * - strlock acquires the sd_lock and sd_reflock to prevent putnext,
4600 * canputnext, backenable etc from dereferencing the q_next that will
4601 * soon change.
4602 * - strlock waits for sd_refcnt to be zero to wait for e.g. any canputnext
4603 * or other q_next walker that uses claimstr/releasestr to finish.
4604 * - optionally for every syncq in the stream strlock acquires all the
4605 * sq_lock's and waits for all sq_counts to drop to a value that indicates
4606 * that no thread executes in the put or service procedures and that no
4607 * thread is draining into the module/driver. This ensures that no
4608 * open, close, put, service, or qtimeout/qbufcall callback procedure is
4609 * currently executing hence no such thread can end up with the old stale
4610 * q_next value and no canput/backenable can have the old stale
4611 * q_nfsrv/q_next.
4612 * - qdetach (wait_svc) makes sure that any scheduled or running threads
4613 * have either finished or observed the QWCLOSE flag and gone away.
4614 */
4615
4616
4617 /*
4618 * Get all the locks necessary to change q_next.
4619 *
4620 * Wait for sd_refcnt to reach 0 and, if sqlist is present, wait for the
4621 * sq_count of each syncq in the list to drop to sq_rmqcount, indicating that
4622 * the only threads inside the syncq are threads currently calling removeq().
4623 * Since threads calling removeq() are in the process of removing their queues
4624 * from the stream, we do not need to worry about them accessing a stale q_next
4625 * pointer and thus we do not need to wait for them to exit (in fact, waiting
4626 * for them can cause deadlock).
4627 *
4628 * This routine is subject to starvation since it does not set any flag to
4629 * prevent threads from entering a module in the stream (i.e. sq_count can
4630 * increase on some syncq while it is waiting on some other syncq).
4631 *
4632 * Assumes that only one thread attempts to call strlock for a given
4633 * stream. If this is not the case the two threads would deadlock.
4634 * This assumption is guaranteed since strlock is only called by insertq
4635 * and removeq and streams plumbing changes are single-threaded for
4636 * a given stream using the STWOPEN, STRCLOSE, and STRPLUMB flags.
4637 *
4638 * For pipes, it is not difficult to atomically designate a pair of streams
4639 * to be mated. Once mated atomically by the framework the twisted pair remain
4640 * configured that way until dismantled atomically by the framework.
4641 * When plumbing takes place on a twisted stream it is necessary to ensure that
4642 * this operation is done exclusively on the twisted stream since two such
4643 * operations, each initiated on different ends of the pipe will deadlock
4644 * waiting for each other to complete.
4645 *
4646 * On entry, no locks should be held.
4647 * The locks acquired and held by strlock depends on a few factors.
4648 * - If sqlist is non-NULL all the syncq locks in the sqlist will be acquired
4649 * and held on exit and all sq_count are at an acceptable level.
4650 * - In all cases, sd_lock and sd_reflock are acquired and held on exit with
4651 * sd_refcnt being zero.
4652 */
4653
4654 static void
4655 strlock(struct stdata *stp, sqlist_t *sqlist)
4656 {
4657 syncql_t *sql, *sql2;
4658 retry:
4659 /*
4660 * Wait for any claimstr to go away.
4661 */
4662 if (STRMATED(stp)) {
4663 struct stdata *stp1, *stp2;
4664
4665 STRLOCKMATES(stp);
4666 /*
4667 * Note that the selection of locking order is not
4668 * important, just that they are always acquired in
4669 * the same order. To assure this, we choose this
4670 * order based on the value of the pointer, and since
4671 * the pointer will not change for the life of this
4672 * pair, we will always grab the locks in the same
4673 * order (and hence, prevent deadlocks).
4674 */
4675 if (&(stp->sd_lock) > &((stp->sd_mate)->sd_lock)) {
4676 stp1 = stp;
4677 stp2 = stp->sd_mate;
4678 } else {
4679 stp2 = stp;
4680 stp1 = stp->sd_mate;
4681 }
4682 mutex_enter(&stp1->sd_reflock);
4683 if (stp1->sd_refcnt > 0) {
4684 STRUNLOCKMATES(stp);
4685 cv_wait(&stp1->sd_refmonitor, &stp1->sd_reflock);
4686 mutex_exit(&stp1->sd_reflock);
4687 goto retry;
4688 }
4689 mutex_enter(&stp2->sd_reflock);
4690 if (stp2->sd_refcnt > 0) {
4691 STRUNLOCKMATES(stp);
4692 mutex_exit(&stp1->sd_reflock);
4693 cv_wait(&stp2->sd_refmonitor, &stp2->sd_reflock);
4694 mutex_exit(&stp2->sd_reflock);
4695 goto retry;
4696 }
4697 STREAM_PUTLOCKS_ENTER(stp1);
4698 STREAM_PUTLOCKS_ENTER(stp2);
4699 } else {
4700 mutex_enter(&stp->sd_lock);
4701 mutex_enter(&stp->sd_reflock);
4702 while (stp->sd_refcnt > 0) {
4703 mutex_exit(&stp->sd_lock);
4704 cv_wait(&stp->sd_refmonitor, &stp->sd_reflock);
4705 if (mutex_tryenter(&stp->sd_lock) == 0) {
4706 mutex_exit(&stp->sd_reflock);
4707 mutex_enter(&stp->sd_lock);
4708 mutex_enter(&stp->sd_reflock);
4709 }
4710 }
4711 STREAM_PUTLOCKS_ENTER(stp);
4712 }
4713
4714 if (sqlist == NULL)
4715 return;
4716
4717 for (sql = sqlist->sqlist_head; sql; sql = sql->sql_next) {
4718 syncq_t *sq = sql->sql_sq;
4719 uint16_t count;
4720
4721 mutex_enter(SQLOCK(sq));
4722 count = sq->sq_count;
4723 ASSERT(sq->sq_rmqcount <= count);
4724 SQ_PUTLOCKS_ENTER(sq);
4725 SUM_SQ_PUTCOUNTS(sq, count);
4726 if (count == sq->sq_rmqcount)
4727 continue;
4728
4729 /* Failed - drop all locks that we have acquired so far */
4730 if (STRMATED(stp)) {
4731 STREAM_PUTLOCKS_EXIT(stp);
4732 STREAM_PUTLOCKS_EXIT(stp->sd_mate);
4733 STRUNLOCKMATES(stp);
4734 mutex_exit(&stp->sd_reflock);
4735 mutex_exit(&stp->sd_mate->sd_reflock);
4736 } else {
4737 STREAM_PUTLOCKS_EXIT(stp);
4738 mutex_exit(&stp->sd_lock);
4739 mutex_exit(&stp->sd_reflock);
4740 }
4741 for (sql2 = sqlist->sqlist_head; sql2 != sql;
4742 sql2 = sql2->sql_next) {
4743 SQ_PUTLOCKS_EXIT(sql2->sql_sq);
4744 mutex_exit(SQLOCK(sql2->sql_sq));
4745 }
4746
4747 /*
4748 * The wait loop below may starve when there are many threads
4749 * claiming the syncq. This is especially a problem with permod
4750 * syncqs (IP). To lessen the impact of the problem we increment
4751 * sq_needexcl and clear fastbits so that putnexts will slow
4752 * down and call sqenable instead of draining right away.
4753 */
4754 sq->sq_needexcl++;
4755 SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
4756 while (count > sq->sq_rmqcount) {
4757 sq->sq_flags |= SQ_WANTWAKEUP;
4758 SQ_PUTLOCKS_EXIT(sq);
4759 cv_wait(&sq->sq_wait, SQLOCK(sq));
4760 count = sq->sq_count;
4761 SQ_PUTLOCKS_ENTER(sq);
4762 SUM_SQ_PUTCOUNTS(sq, count);
4763 }
4764 sq->sq_needexcl--;
4765 if (sq->sq_needexcl == 0)
4766 SQ_PUTCOUNT_SETFAST_LOCKED(sq);
4767 SQ_PUTLOCKS_EXIT(sq);
4768 ASSERT(count == sq->sq_rmqcount);
4769 mutex_exit(SQLOCK(sq));
4770 goto retry;
4771 }
4772 }
4773
4774 /*
4775 * Drop all the locks that strlock acquired.
4776 */
4777 static void
4778 strunlock(struct stdata *stp, sqlist_t *sqlist)
4779 {
4780 syncql_t *sql;
4781
4782 if (STRMATED(stp)) {
4783 STREAM_PUTLOCKS_EXIT(stp);
4784 STREAM_PUTLOCKS_EXIT(stp->sd_mate);
4785 STRUNLOCKMATES(stp);
4786 mutex_exit(&stp->sd_reflock);
4787 mutex_exit(&stp->sd_mate->sd_reflock);
4788 } else {
4789 STREAM_PUTLOCKS_EXIT(stp);
4790 mutex_exit(&stp->sd_lock);
4791 mutex_exit(&stp->sd_reflock);
4792 }
4793
4794 if (sqlist == NULL)
4795 return;
4796
4797 for (sql = sqlist->sqlist_head; sql; sql = sql->sql_next) {
4798 SQ_PUTLOCKS_EXIT(sql->sql_sq);
4799 mutex_exit(SQLOCK(sql->sql_sq));
4800 }
4801 }
4802
4803 /*
4804 * When the module has service procedure, we need check if the next
4805 * module which has service procedure is in flow control to trigger
4806 * the backenable.
4807 */
4808 static void
4809 backenable_insertedq(queue_t *q)
4810 {
4811 qband_t *qbp;
4812
4813 claimstr(q);
4814 if (q->q_qinfo->qi_srvp != NULL && q->q_next != NULL) {
4815 if (q->q_next->q_nfsrv->q_flag & QWANTW)
4816 backenable(q, 0);
4817
4818 qbp = q->q_next->q_nfsrv->q_bandp;
4819 for (; qbp != NULL; qbp = qbp->qb_next)
4820 if ((qbp->qb_flag & QB_WANTW) && qbp->qb_first != NULL)
4821 backenable(q, qbp->qb_first->b_band);
4822 }
4823 releasestr(q);
4824 }
4825
4826 /*
4827 * Given two read queues, insert a new single one after another.
4828 *
4829 * This routine acquires all the necessary locks in order to change
4830 * q_next and related pointer using strlock().
4831 * It depends on the stream head ensuring that there are no concurrent
4832 * insertq or removeq on the same stream. The stream head ensures this
4833 * using the flags STWOPEN, STRCLOSE, and STRPLUMB.
4834 *
4835 * Note that no syncq locks are held during the q_next change. This is
4836 * applied to all streams since, unlike removeq, there is no problem of stale
4837 * pointers when adding a module to the stream. Thus drivers/modules that do a
4838 * canput(rq->q_next) would never get a closed/freed queue pointer even if we
4839 * applied this optimization to all streams.
4840 */
4841 void
4842 insertq(struct stdata *stp, queue_t *new)
4843 {
4844 queue_t *after;
4845 queue_t *wafter;
4846 queue_t *wnew = _WR(new);
4847 boolean_t have_fifo = B_FALSE;
4848
4849 if (new->q_flag & _QINSERTING) {
4850 ASSERT(stp->sd_vnode->v_type != VFIFO);
4851 after = new->q_next;
4852 wafter = _WR(new->q_next);
4853 } else {
4854 after = _RD(stp->sd_wrq);
4855 wafter = stp->sd_wrq;
4856 }
4857
4858 TRACE_2(TR_FAC_STREAMS_FR, TR_INSERTQ,
4859 "insertq:%p, %p", after, new);
4860 ASSERT(after->q_flag & QREADR);
4861 ASSERT(new->q_flag & QREADR);
4862
4863 strlock(stp, NULL);
4864
4865 /* Do we have a FIFO? */
4866 if (wafter->q_next == after) {
4867 have_fifo = B_TRUE;
4868 wnew->q_next = new;
4869 } else {
4870 wnew->q_next = wafter->q_next;
4871 }
4872 new->q_next = after;
4873
4874 set_nfsrv_ptr(new, wnew, after, wafter);
4875 /*
4876 * set_nfsrv_ptr() needs to know if this is an insertion or not,
4877 * so only reset this flag after calling it.
4878 */
4879 new->q_flag &= ~_QINSERTING;
4880
4881 if (have_fifo) {
4882 wafter->q_next = wnew;
4883 } else {
4884 if (wafter->q_next)
4885 _OTHERQ(wafter->q_next)->q_next = new;
4886 wafter->q_next = wnew;
4887 }
4888
4889 set_qend(new);
4890 /* The QEND flag might have to be updated for the upstream guy */
4891 set_qend(after);
4892
4893 ASSERT(_SAMESTR(new) == O_SAMESTR(new));
4894 ASSERT(_SAMESTR(wnew) == O_SAMESTR(wnew));
4895 ASSERT(_SAMESTR(after) == O_SAMESTR(after));
4896 ASSERT(_SAMESTR(wafter) == O_SAMESTR(wafter));
4897 strsetuio(stp);
4898
4899 /*
4900 * If this was a module insertion, bump the push count.
4901 */
4902 if (!(new->q_flag & QISDRV))
4903 stp->sd_pushcnt++;
4904
4905 strunlock(stp, NULL);
4906
4907 /* check if the write Q needs backenable */
4908 backenable_insertedq(wnew);
4909
4910 /* check if the read Q needs backenable */
4911 backenable_insertedq(new);
4912 }
4913
4914 /*
4915 * Given a read queue, unlink it from any neighbors.
4916 *
4917 * This routine acquires all the necessary locks in order to
4918 * change q_next and related pointers and also guard against
4919 * stale references (e.g. through q_next) to the queue that
4920 * is being removed. It also plays part of the role in ensuring
4921 * that the module's/driver's put procedure doesn't get called
4922 * after qprocsoff returns.
4923 *
4924 * Removeq depends on the stream head ensuring that there are
4925 * no concurrent insertq or removeq on the same stream. The
4926 * stream head ensures this using the flags STWOPEN, STRCLOSE and
4927 * STRPLUMB.
4928 *
4929 * The set of locks needed to remove the queue is different in
4930 * different cases:
4931 *
4932 * Acquire sd_lock, sd_reflock, and all the syncq locks in the stream after
4933 * waiting for the syncq reference count to drop to 0 indicating that no
4934 * non-close threads are present anywhere in the stream. This ensures that any
4935 * module/driver can reference q_next in its open, close, put, or service
4936 * procedures.
4937 *
4938 * The sq_rmqcount counter tracks the number of threads inside removeq().
4939 * strlock() ensures that there is either no threads executing inside perimeter
4940 * or there is only a thread calling qprocsoff().
4941 *
4942 * strlock() compares the value of sq_count with the number of threads inside
4943 * removeq() and waits until sq_count is equal to sq_rmqcount. We need to wakeup
4944 * any threads waiting in strlock() when the sq_rmqcount increases.
4945 */
4946
4947 void
4948 removeq(queue_t *qp)
4949 {
4950 queue_t *wqp = _WR(qp);
4951 struct stdata *stp = STREAM(qp);
4952 sqlist_t *sqlist = NULL;
4953 boolean_t isdriver;
4954 int moved;
4955 syncq_t *sq = qp->q_syncq;
4956 syncq_t *wsq = wqp->q_syncq;
4957
4958 ASSERT(stp);
4959
4960 TRACE_2(TR_FAC_STREAMS_FR, TR_REMOVEQ,
4961 "removeq:%p %p", qp, wqp);
4962 ASSERT(qp->q_flag&QREADR);
4963
4964 /*
4965 * For queues using Synchronous streams, we must wait for all threads in
4966 * rwnext() to drain out before proceeding.
4967 */
4968 if (qp->q_flag & QSYNCSTR) {
4969 /* First, we need wakeup any threads blocked in rwnext() */
4970 mutex_enter(SQLOCK(sq));
4971 if (sq->sq_flags & SQ_WANTWAKEUP) {
4972 sq->sq_flags &= ~SQ_WANTWAKEUP;
4973 cv_broadcast(&sq->sq_wait);
4974 }
4975 mutex_exit(SQLOCK(sq));
4976
4977 if (wsq != sq) {
4978 mutex_enter(SQLOCK(wsq));
4979 if (wsq->sq_flags & SQ_WANTWAKEUP) {
4980 wsq->sq_flags &= ~SQ_WANTWAKEUP;
4981 cv_broadcast(&wsq->sq_wait);
4982 }
4983 mutex_exit(SQLOCK(wsq));
4984 }
4985
4986 mutex_enter(QLOCK(qp));
4987 while (qp->q_rwcnt > 0) {
4988 qp->q_flag |= QWANTRMQSYNC;
4989 cv_wait(&qp->q_wait, QLOCK(qp));
4990 }
4991 mutex_exit(QLOCK(qp));
4992
4993 mutex_enter(QLOCK(wqp));
4994 while (wqp->q_rwcnt > 0) {
4995 wqp->q_flag |= QWANTRMQSYNC;
4996 cv_wait(&wqp->q_wait, QLOCK(wqp));
4997 }
4998 mutex_exit(QLOCK(wqp));
4999 }
5000
5001 mutex_enter(SQLOCK(sq));
5002 sq->sq_rmqcount++;
5003 if (sq->sq_flags & SQ_WANTWAKEUP) {
5004 sq->sq_flags &= ~SQ_WANTWAKEUP;
5005 cv_broadcast(&sq->sq_wait);
5006 }
5007 mutex_exit(SQLOCK(sq));
5008
5009 isdriver = (qp->q_flag & QISDRV);
5010
5011 sqlist = sqlist_build(qp, stp, STRMATED(stp));
5012 strlock(stp, sqlist);
5013
5014 reset_nfsrv_ptr(qp, wqp);
5015
5016 ASSERT(wqp->q_next == NULL || backq(qp)->q_next == qp);
5017 ASSERT(qp->q_next == NULL || backq(wqp)->q_next == wqp);
5018 /* Do we have a FIFO? */
5019 if (wqp->q_next == qp) {
5020 stp->sd_wrq->q_next = _RD(stp->sd_wrq);
5021 } else {
5022 if (wqp->q_next)
5023 backq(qp)->q_next = qp->q_next;
5024 if (qp->q_next)
5025 backq(wqp)->q_next = wqp->q_next;
5026 }
5027
5028 /* The QEND flag might have to be updated for the upstream guy */
5029 if (qp->q_next)
5030 set_qend(qp->q_next);
5031
5032 ASSERT(_SAMESTR(stp->sd_wrq) == O_SAMESTR(stp->sd_wrq));
5033 ASSERT(_SAMESTR(_RD(stp->sd_wrq)) == O_SAMESTR(_RD(stp->sd_wrq)));
5034
5035 /*
5036 * Move any messages destined for the put procedures to the next
5037 * syncq in line. Otherwise free them.
5038 */
5039 moved = 0;
5040 /*
5041 * Quick check to see whether there are any messages or events.
5042 */
5043 if (qp->q_syncqmsgs != 0 || (qp->q_syncq->sq_flags & SQ_EVENTS))
5044 moved += propagate_syncq(qp);
5045 if (wqp->q_syncqmsgs != 0 ||
5046 (wqp->q_syncq->sq_flags & SQ_EVENTS))
5047 moved += propagate_syncq(wqp);
5048
5049 strsetuio(stp);
5050
5051 /*
5052 * If this was a module removal, decrement the push count.
5053 */
5054 if (!isdriver)
5055 stp->sd_pushcnt--;
5056
5057 strunlock(stp, sqlist);
5058 sqlist_free(sqlist);
5059
5060 /*
5061 * Make sure any messages that were propagated are drained.
5062 * Also clear any QFULL bit caused by messages that were propagated.
5063 */
5064
5065 if (qp->q_next != NULL) {
5066 clr_qfull(qp);
5067 /*
5068 * For the driver calling qprocsoff, propagate_syncq
5069 * frees all the messages instead of putting it in
5070 * the stream head
5071 */
5072 if (!isdriver && (moved > 0))
5073 emptysq(qp->q_next->q_syncq);
5074 }
5075 if (wqp->q_next != NULL) {
5076 clr_qfull(wqp);
5077 /*
5078 * We come here for any pop of a module except for the
5079 * case of driver being removed. We don't call emptysq
5080 * if we did not move any messages. This will avoid holding
5081 * PERMOD syncq locks in emptysq
5082 */
5083 if (moved > 0)
5084 emptysq(wqp->q_next->q_syncq);
5085 }
5086
5087 mutex_enter(SQLOCK(sq));
5088 sq->sq_rmqcount--;
5089 mutex_exit(SQLOCK(sq));
5090 }
5091
5092 /*
5093 * Prevent further entry by setting a flag (like SQ_FROZEN, SQ_BLOCKED or
5094 * SQ_WRITER) on a syncq.
5095 * If maxcnt is not -1 it assumes that caller has "maxcnt" claim(s) on the
5096 * sync queue and waits until sq_count reaches maxcnt.
5097 *
5098 * If maxcnt is -1 there's no need to grab sq_putlocks since the caller
5099 * does not care about putnext threads that are in the middle of calling put
5100 * entry points.
5101 *
5102 * This routine is used for both inner and outer syncqs.
5103 */
5104 static void
5105 blocksq(syncq_t *sq, ushort_t flag, int maxcnt)
5106 {
5107 uint16_t count = 0;
5108
5109 mutex_enter(SQLOCK(sq));
5110 /*
5111 * Wait for SQ_FROZEN/SQ_BLOCKED to be reset.
5112 * SQ_FROZEN will be set if there is a frozen stream that has a
5113 * queue which also refers to this "shared" syncq.
5114 * SQ_BLOCKED will be set if there is "off" queue which also
5115 * refers to this "shared" syncq.
5116 */
5117 if (maxcnt != -1) {
5118 count = sq->sq_count;
5119 SQ_PUTLOCKS_ENTER(sq);
5120 SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
5121 SUM_SQ_PUTCOUNTS(sq, count);
5122 }
5123 sq->sq_needexcl++;
5124 ASSERT(sq->sq_needexcl != 0); /* wraparound */
5125
5126 while ((sq->sq_flags & flag) ||
5127 (maxcnt != -1 && count > (unsigned)maxcnt)) {
5128 sq->sq_flags |= SQ_WANTWAKEUP;
5129 if (maxcnt != -1) {
5130 SQ_PUTLOCKS_EXIT(sq);
5131 }
5132 cv_wait(&sq->sq_wait, SQLOCK(sq));
5133 if (maxcnt != -1) {
5134 count = sq->sq_count;
5135 SQ_PUTLOCKS_ENTER(sq);
5136 SUM_SQ_PUTCOUNTS(sq, count);
5137 }
5138 }
5139 sq->sq_needexcl--;
5140 sq->sq_flags |= flag;
5141 ASSERT(maxcnt == -1 || count == maxcnt);
5142 if (maxcnt != -1) {
5143 if (sq->sq_needexcl == 0) {
5144 SQ_PUTCOUNT_SETFAST_LOCKED(sq);
5145 }
5146 SQ_PUTLOCKS_EXIT(sq);
5147 } else if (sq->sq_needexcl == 0) {
5148 SQ_PUTCOUNT_SETFAST(sq);
5149 }
5150
5151 mutex_exit(SQLOCK(sq));
5152 }
5153
5154 /*
5155 * Reset a flag that was set with blocksq.
5156 *
5157 * Can not use this routine to reset SQ_WRITER.
5158 *
5159 * If "isouter" is set then the syncq is assumed to be an outer perimeter
5160 * and drain_syncq is not called. Instead we rely on the qwriter_outer thread
5161 * to handle the queued qwriter operations.
5162 *
5163 * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5164 * sq_putlocks are used.
5165 */
5166 static void
5167 unblocksq(syncq_t *sq, uint16_t resetflag, int isouter)
5168 {
5169 uint16_t flags;
5170
5171 mutex_enter(SQLOCK(sq));
5172 ASSERT(resetflag != SQ_WRITER);
5173 ASSERT(sq->sq_flags & resetflag);
5174 flags = sq->sq_flags & ~resetflag;
5175 sq->sq_flags = flags;
5176 if (flags & (SQ_QUEUED | SQ_WANTWAKEUP)) {
5177 if (flags & SQ_WANTWAKEUP) {
5178 flags &= ~SQ_WANTWAKEUP;
5179 cv_broadcast(&sq->sq_wait);
5180 }
5181 sq->sq_flags = flags;
5182 if ((flags & SQ_QUEUED) && !(flags & (SQ_STAYAWAY|SQ_EXCL))) {
5183 if (!isouter) {
5184 /* drain_syncq drops SQLOCK */
5185 drain_syncq(sq);
5186 return;
5187 }
5188 }
5189 }
5190 mutex_exit(SQLOCK(sq));
5191 }
5192
5193 /*
5194 * Reset a flag that was set with blocksq.
5195 * Does not drain the syncq. Use emptysq() for that.
5196 * Returns 1 if SQ_QUEUED is set. Otherwise 0.
5197 *
5198 * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5199 * sq_putlocks are used.
5200 */
5201 static int
5202 dropsq(syncq_t *sq, uint16_t resetflag)
5203 {
5204 uint16_t flags;
5205
5206 mutex_enter(SQLOCK(sq));
5207 ASSERT(sq->sq_flags & resetflag);
5208 flags = sq->sq_flags & ~resetflag;
5209 if (flags & SQ_WANTWAKEUP) {
5210 flags &= ~SQ_WANTWAKEUP;
5211 cv_broadcast(&sq->sq_wait);
5212 }
5213 sq->sq_flags = flags;
5214 mutex_exit(SQLOCK(sq));
5215 if (flags & SQ_QUEUED)
5216 return (1);
5217 return (0);
5218 }
5219
5220 /*
5221 * Empty all the messages on a syncq.
5222 *
5223 * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5224 * sq_putlocks are used.
5225 */
5226 static void
5227 emptysq(syncq_t *sq)
5228 {
5229 uint16_t flags;
5230
5231 mutex_enter(SQLOCK(sq));
5232 flags = sq->sq_flags;
5233 if ((flags & SQ_QUEUED) && !(flags & (SQ_STAYAWAY|SQ_EXCL))) {
5234 /*
5235 * To prevent potential recursive invocation of drain_syncq we
5236 * do not call drain_syncq if count is non-zero.
5237 */
5238 if (sq->sq_count == 0) {
5239 /* drain_syncq() drops SQLOCK */
5240 drain_syncq(sq);
5241 return;
5242 } else
5243 sqenable(sq);
5244 }
5245 mutex_exit(SQLOCK(sq));
5246 }
5247
5248 /*
5249 * Ordered insert while removing duplicates.
5250 */
5251 static void
5252 sqlist_insert(sqlist_t *sqlist, syncq_t *sqp)
5253 {
5254 syncql_t *sqlp, **prev_sqlpp, *new_sqlp;
5255
5256 prev_sqlpp = &sqlist->sqlist_head;
5257 while ((sqlp = *prev_sqlpp) != NULL) {
5258 if (sqlp->sql_sq >= sqp) {
5259 if (sqlp->sql_sq == sqp) /* duplicate */
5260 return;
5261 break;
5262 }
5263 prev_sqlpp = &sqlp->sql_next;
5264 }
5265 new_sqlp = &sqlist->sqlist_array[sqlist->sqlist_index++];
5266 ASSERT((char *)new_sqlp < (char *)sqlist + sqlist->sqlist_size);
5267 new_sqlp->sql_next = sqlp;
5268 new_sqlp->sql_sq = sqp;
5269 *prev_sqlpp = new_sqlp;
5270 }
5271
5272 /*
5273 * Walk the write side queues until we hit either the driver
5274 * or a twist in the stream (_SAMESTR will return false in both
5275 * these cases) then turn around and walk the read side queues
5276 * back up to the stream head.
5277 */
5278 static void
5279 sqlist_insertall(sqlist_t *sqlist, queue_t *q)
5280 {
5281 while (q != NULL) {
5282 sqlist_insert(sqlist, q->q_syncq);
5283
5284 if (_SAMESTR(q))
5285 q = q->q_next;
5286 else if (!(q->q_flag & QREADR))
5287 q = _RD(q);
5288 else
5289 q = NULL;
5290 }
5291 }
5292
5293 /*
5294 * Allocate and build a list of all syncqs in a stream and the syncq(s)
5295 * associated with the "q" parameter. The resulting list is sorted in a
5296 * canonical order and is free of duplicates.
5297 * Assumes the passed queue is a _RD(q).
5298 */
5299 static sqlist_t *
5300 sqlist_build(queue_t *q, struct stdata *stp, boolean_t do_twist)
5301 {
5302 sqlist_t *sqlist = sqlist_alloc(stp, KM_SLEEP);
5303
5304 /*
5305 * start with the current queue/qpair
5306 */
5307 ASSERT(q->q_flag & QREADR);
5308
5309 sqlist_insert(sqlist, q->q_syncq);
5310 sqlist_insert(sqlist, _WR(q)->q_syncq);
5311
5312 sqlist_insertall(sqlist, stp->sd_wrq);
5313 if (do_twist)
5314 sqlist_insertall(sqlist, stp->sd_mate->sd_wrq);
5315
5316 return (sqlist);
5317 }
5318
5319 static sqlist_t *
5320 sqlist_alloc(struct stdata *stp, int kmflag)
5321 {
5322 size_t sqlist_size;
5323 sqlist_t *sqlist;
5324
5325 /*
5326 * Allocate 2 syncql_t's for each pushed module. Note that
5327 * the sqlist_t structure already has 4 syncql_t's built in:
5328 * 2 for the stream head, and 2 for the driver/other stream head.
5329 */
5330 sqlist_size = 2 * sizeof (syncql_t) * stp->sd_pushcnt +
5331 sizeof (sqlist_t);
5332 if (STRMATED(stp))
5333 sqlist_size += 2 * sizeof (syncql_t) * stp->sd_mate->sd_pushcnt;
5334 sqlist = kmem_alloc(sqlist_size, kmflag);
5335
5336 sqlist->sqlist_head = NULL;
5337 sqlist->sqlist_size = sqlist_size;
5338 sqlist->sqlist_index = 0;
5339
5340 return (sqlist);
5341 }
5342
5343 /*
5344 * Free the list created by sqlist_alloc()
5345 */
5346 static void
5347 sqlist_free(sqlist_t *sqlist)
5348 {
5349 kmem_free(sqlist, sqlist->sqlist_size);
5350 }
5351
5352 /*
5353 * Prevent any new entries into any syncq in this stream.
5354 * Used by freezestr.
5355 */
5356 void
5357 strblock(queue_t *q)
5358 {
5359 struct stdata *stp;
5360 syncql_t *sql;
5361 sqlist_t *sqlist;
5362
5363 q = _RD(q);
5364
5365 stp = STREAM(q);
5366 ASSERT(stp != NULL);
5367
5368 /*
5369 * Get a sorted list with all the duplicates removed containing
5370 * all the syncqs referenced by this stream.
5371 */
5372 sqlist = sqlist_build(q, stp, B_FALSE);
5373 for (sql = sqlist->sqlist_head; sql != NULL; sql = sql->sql_next)
5374 blocksq(sql->sql_sq, SQ_FROZEN, -1);
5375 sqlist_free(sqlist);
5376 }
5377
5378 /*
5379 * Release the block on new entries into this stream
5380 */
5381 void
5382 strunblock(queue_t *q)
5383 {
5384 struct stdata *stp;
5385 syncql_t *sql;
5386 sqlist_t *sqlist;
5387 int drain_needed;
5388
5389 q = _RD(q);
5390
5391 /*
5392 * Get a sorted list with all the duplicates removed containing
5393 * all the syncqs referenced by this stream.
5394 * Have to drop the SQ_FROZEN flag on all the syncqs before
5395 * starting to drain them; otherwise the draining might
5396 * cause a freezestr in some module on the stream (which
5397 * would deadlock).
5398 */
5399 stp = STREAM(q);
5400 ASSERT(stp != NULL);
5401 sqlist = sqlist_build(q, stp, B_FALSE);
5402 drain_needed = 0;
5403 for (sql = sqlist->sqlist_head; sql != NULL; sql = sql->sql_next)
5404 drain_needed += dropsq(sql->sql_sq, SQ_FROZEN);
5405 if (drain_needed) {
5406 for (sql = sqlist->sqlist_head; sql != NULL;
5407 sql = sql->sql_next)
5408 emptysq(sql->sql_sq);
5409 }
5410 sqlist_free(sqlist);
5411 }
5412
5413 #ifdef DEBUG
5414 static int
5415 qprocsareon(queue_t *rq)
5416 {
5417 if (rq->q_next == NULL)
5418 return (0);
5419 return (_WR(rq->q_next)->q_next == _WR(rq));
5420 }
5421
5422 int
5423 qclaimed(queue_t *q)
5424 {
5425 uint_t count;
5426
5427 count = q->q_syncq->sq_count;
5428 SUM_SQ_PUTCOUNTS(q->q_syncq, count);
5429 return (count != 0);
5430 }
5431
5432 /*
5433 * Check if anyone has frozen this stream with freezestr
5434 */
5435 int
5436 frozenstr(queue_t *q)
5437 {
5438 return ((q->q_syncq->sq_flags & SQ_FROZEN) != 0);
5439 }
5440 #endif /* DEBUG */
5441
5442 /*
5443 * Enter a queue.
5444 * Obsoleted interface. Should not be used.
5445 */
5446 void
5447 enterq(queue_t *q)
5448 {
5449 entersq(q->q_syncq, SQ_CALLBACK);
5450 }
5451
5452 void
5453 leaveq(queue_t *q)
5454 {
5455 leavesq(q->q_syncq, SQ_CALLBACK);
5456 }
5457
5458 /*
5459 * Enter a perimeter. c_inner and c_outer specifies which concurrency bits
5460 * to check.
5461 * Wait if SQ_QUEUED is set to preserve ordering between messages and qwriter
5462 * calls and the running of open, close and service procedures.
5463 *
5464 * If c_inner bit is set no need to grab sq_putlocks since we don't care
5465 * if other threads have entered or are entering put entry point.
5466 *
5467 * If c_inner bit is set it might have been possible to use
5468 * sq_putlocks/sq_putcounts instead of SQLOCK/sq_count (e.g. to optimize
5469 * open/close path for IP) but since the count may need to be decremented in
5470 * qwait() we wouldn't know which counter to decrement. Currently counter is
5471 * selected by current cpu_seqid and current CPU can change at any moment. XXX
5472 * in the future we might use curthread id bits to select the counter and this
5473 * would stay constant across routine calls.
5474 */
5475 void
5476 entersq(syncq_t *sq, int entrypoint)
5477 {
5478 uint16_t count = 0;
5479 uint16_t flags;
5480 uint16_t waitflags = SQ_STAYAWAY | SQ_EVENTS | SQ_EXCL;
5481 uint16_t type;
5482 uint_t c_inner = entrypoint & SQ_CI;
5483 uint_t c_outer = entrypoint & SQ_CO;
5484
5485 /*
5486 * Increment ref count to keep closes out of this queue.
5487 */
5488 ASSERT(sq);
5489 ASSERT(c_inner && c_outer);
5490 mutex_enter(SQLOCK(sq));
5491 flags = sq->sq_flags;
5492 type = sq->sq_type;
5493 if (!(type & c_inner)) {
5494 /* Make sure all putcounts now use slowlock. */
5495 count = sq->sq_count;
5496 SQ_PUTLOCKS_ENTER(sq);
5497 SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
5498 SUM_SQ_PUTCOUNTS(sq, count);
5499 sq->sq_needexcl++;
5500 ASSERT(sq->sq_needexcl != 0); /* wraparound */
5501 waitflags |= SQ_MESSAGES;
5502 }
5503 /*
5504 * Wait until we can enter the inner perimeter.
5505 * If we want exclusive access we wait until sq_count is 0.
5506 * We have to do this before entering the outer perimeter in order
5507 * to preserve put/close message ordering.
5508 */
5509 while ((flags & waitflags) || (!(type & c_inner) && count != 0)) {
5510 sq->sq_flags = flags | SQ_WANTWAKEUP;
5511 if (!(type & c_inner)) {
5512 SQ_PUTLOCKS_EXIT(sq);
5513 }
5514 cv_wait(&sq->sq_wait, SQLOCK(sq));
5515 if (!(type & c_inner)) {
5516 count = sq->sq_count;
5517 SQ_PUTLOCKS_ENTER(sq);
5518 SUM_SQ_PUTCOUNTS(sq, count);
5519 }
5520 flags = sq->sq_flags;
5521 }
5522
5523 if (!(type & c_inner)) {
5524 ASSERT(sq->sq_needexcl > 0);
5525 sq->sq_needexcl--;
5526 if (sq->sq_needexcl == 0) {
5527 SQ_PUTCOUNT_SETFAST_LOCKED(sq);
5528 }
5529 }
5530
5531 /* Check if we need to enter the outer perimeter */
5532 if (!(type & c_outer)) {
5533 /*
5534 * We have to enter the outer perimeter exclusively before
5535 * we can increment sq_count to avoid deadlock. This implies
5536 * that we have to re-check sq_flags and sq_count.
5537 *
5538 * is it possible to have c_inner set when c_outer is not set?
5539 */
5540 if (!(type & c_inner)) {
5541 SQ_PUTLOCKS_EXIT(sq);
5542 }
5543 mutex_exit(SQLOCK(sq));
5544 outer_enter(sq->sq_outer, SQ_GOAWAY);
5545 mutex_enter(SQLOCK(sq));
5546 flags = sq->sq_flags;
5547 /*
5548 * there should be no need to recheck sq_putcounts
5549 * because outer_enter() has already waited for them to clear
5550 * after setting SQ_WRITER.
5551 */
5552 count = sq->sq_count;
5553 #ifdef DEBUG
5554 /*
5555 * SUMCHECK_SQ_PUTCOUNTS should return the sum instead
5556 * of doing an ASSERT internally. Others should do
5557 * something like
5558 * ASSERT(SUMCHECK_SQ_PUTCOUNTS(sq) == 0);
5559 * without the need to #ifdef DEBUG it.
5560 */
5561 SUMCHECK_SQ_PUTCOUNTS(sq, 0);
5562 #endif
5563 while ((flags & (SQ_EXCL|SQ_BLOCKED|SQ_FROZEN)) ||
5564 (!(type & c_inner) && count != 0)) {
5565 sq->sq_flags = flags | SQ_WANTWAKEUP;
5566 cv_wait(&sq->sq_wait, SQLOCK(sq));
5567 count = sq->sq_count;
5568 flags = sq->sq_flags;
5569 }
5570 }
5571
5572 sq->sq_count++;
5573 ASSERT(sq->sq_count != 0); /* Wraparound */
5574 if (!(type & c_inner)) {
5575 /* Exclusive entry */
5576 ASSERT(sq->sq_count == 1);
5577 sq->sq_flags |= SQ_EXCL;
5578 if (type & c_outer) {
5579 SQ_PUTLOCKS_EXIT(sq);
5580 }
5581 }
5582 mutex_exit(SQLOCK(sq));
5583 }
5584
5585 /*
5586 * Leave a syncq. Announce to framework that closes may proceed.
5587 * c_inner and c_outer specify which concurrency bits to check.
5588 *
5589 * Must never be called from driver or module put entry point.
5590 *
5591 * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5592 * sq_putlocks are used.
5593 */
5594 void
5595 leavesq(syncq_t *sq, int entrypoint)
5596 {
5597 uint16_t flags;
5598 uint16_t type;
5599 uint_t c_outer = entrypoint & SQ_CO;
5600 #ifdef DEBUG
5601 uint_t c_inner = entrypoint & SQ_CI;
5602 #endif
5603
5604 /*
5605 * Decrement ref count, drain the syncq if possible, and wake up
5606 * any waiting close.
5607 */
5608 ASSERT(sq);
5609 ASSERT(c_inner && c_outer);
5610 mutex_enter(SQLOCK(sq));
5611 flags = sq->sq_flags;
5612 type = sq->sq_type;
5613 if (flags & (SQ_QUEUED|SQ_WANTWAKEUP|SQ_WANTEXWAKEUP)) {
5614
5615 if (flags & SQ_WANTWAKEUP) {
5616 flags &= ~SQ_WANTWAKEUP;
5617 cv_broadcast(&sq->sq_wait);
5618 }
5619 if (flags & SQ_WANTEXWAKEUP) {
5620 flags &= ~SQ_WANTEXWAKEUP;
5621 cv_broadcast(&sq->sq_exitwait);
5622 }
5623
5624 if ((flags & SQ_QUEUED) && !(flags & SQ_STAYAWAY)) {
5625 /*
5626 * The syncq needs to be drained. "Exit" the syncq
5627 * before calling drain_syncq.
5628 */
5629 ASSERT(sq->sq_count != 0);
5630 sq->sq_count--;
5631 ASSERT((flags & SQ_EXCL) || (type & c_inner));
5632 sq->sq_flags = flags & ~SQ_EXCL;
5633 drain_syncq(sq);
5634 ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
5635 /* Check if we need to exit the outer perimeter */
5636 /* XXX will this ever be true? */
5637 if (!(type & c_outer))
5638 outer_exit(sq->sq_outer);
5639 return;
5640 }
5641 }
5642 ASSERT(sq->sq_count != 0);
5643 sq->sq_count--;
5644 ASSERT((flags & SQ_EXCL) || (type & c_inner));
5645 sq->sq_flags = flags & ~SQ_EXCL;
5646 mutex_exit(SQLOCK(sq));
5647
5648 /* Check if we need to exit the outer perimeter */
5649 if (!(sq->sq_type & c_outer))
5650 outer_exit(sq->sq_outer);
5651 }
5652
5653 /*
5654 * Prevent q_next from changing in this stream by incrementing sq_count.
5655 *
5656 * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5657 * sq_putlocks are used.
5658 */
5659 void
5660 claimq(queue_t *qp)
5661 {
5662 syncq_t *sq = qp->q_syncq;
5663
5664 mutex_enter(SQLOCK(sq));
5665 sq->sq_count++;
5666 ASSERT(sq->sq_count != 0); /* Wraparound */
5667 mutex_exit(SQLOCK(sq));
5668 }
5669
5670 /*
5671 * Undo claimq.
5672 *
5673 * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5674 * sq_putlocks are used.
5675 */
5676 void
5677 releaseq(queue_t *qp)
5678 {
5679 syncq_t *sq = qp->q_syncq;
5680 uint16_t flags;
5681
5682 mutex_enter(SQLOCK(sq));
5683 ASSERT(sq->sq_count > 0);
5684 sq->sq_count--;
5685
5686 flags = sq->sq_flags;
5687 if (flags & (SQ_WANTWAKEUP|SQ_QUEUED)) {
5688 if (flags & SQ_WANTWAKEUP) {
5689 flags &= ~SQ_WANTWAKEUP;
5690 cv_broadcast(&sq->sq_wait);
5691 }
5692 sq->sq_flags = flags;
5693 if ((flags & SQ_QUEUED) && !(flags & (SQ_STAYAWAY|SQ_EXCL))) {
5694 /*
5695 * To prevent potential recursive invocation of
5696 * drain_syncq we do not call drain_syncq if count is
5697 * non-zero.
5698 */
5699 if (sq->sq_count == 0) {
5700 drain_syncq(sq);
5701 return;
5702 } else
5703 sqenable(sq);
5704 }
5705 }
5706 mutex_exit(SQLOCK(sq));
5707 }
5708
5709 /*
5710 * Prevent q_next from changing in this stream by incrementing sd_refcnt.
5711 */
5712 void
5713 claimstr(queue_t *qp)
5714 {
5715 struct stdata *stp = STREAM(qp);
5716
5717 mutex_enter(&stp->sd_reflock);
5718 stp->sd_refcnt++;
5719 ASSERT(stp->sd_refcnt != 0); /* Wraparound */
5720 mutex_exit(&stp->sd_reflock);
5721 }
5722
5723 /*
5724 * Undo claimstr.
5725 */
5726 void
5727 releasestr(queue_t *qp)
5728 {
5729 struct stdata *stp = STREAM(qp);
5730
5731 mutex_enter(&stp->sd_reflock);
5732 ASSERT(stp->sd_refcnt != 0);
5733 if (--stp->sd_refcnt == 0)
5734 cv_broadcast(&stp->sd_refmonitor);
5735 mutex_exit(&stp->sd_reflock);
5736 }
5737
5738 static syncq_t *
5739 new_syncq(void)
5740 {
5741 return (kmem_cache_alloc(syncq_cache, KM_SLEEP));
5742 }
5743
5744 static void
5745 free_syncq(syncq_t *sq)
5746 {
5747 ASSERT(sq->sq_head == NULL);
5748 ASSERT(sq->sq_outer == NULL);
5749 ASSERT(sq->sq_callbpend == NULL);
5750 ASSERT((sq->sq_onext == NULL && sq->sq_oprev == NULL) ||
5751 (sq->sq_onext == sq && sq->sq_oprev == sq));
5752
5753 if (sq->sq_ciputctrl != NULL) {
5754 ASSERT(sq->sq_nciputctrl == n_ciputctrl - 1);
5755 SUMCHECK_CIPUTCTRL_COUNTS(sq->sq_ciputctrl,
5756 sq->sq_nciputctrl, 0);
5757 ASSERT(ciputctrl_cache != NULL);
5758 kmem_cache_free(ciputctrl_cache, sq->sq_ciputctrl);
5759 }
5760
5761 sq->sq_tail = NULL;
5762 sq->sq_evhead = NULL;
5763 sq->sq_evtail = NULL;
5764 sq->sq_ciputctrl = NULL;
5765 sq->sq_nciputctrl = 0;
5766 sq->sq_count = 0;
5767 sq->sq_rmqcount = 0;
5768 sq->sq_callbflags = 0;
5769 sq->sq_cancelid = 0;
5770 sq->sq_next = NULL;
5771 sq->sq_needexcl = 0;
5772 sq->sq_svcflags = 0;
5773 sq->sq_nqueues = 0;
5774 sq->sq_pri = 0;
5775 sq->sq_onext = NULL;
5776 sq->sq_oprev = NULL;
5777 sq->sq_flags = 0;
5778 sq->sq_type = 0;
5779 sq->sq_servcount = 0;
5780
5781 kmem_cache_free(syncq_cache, sq);
5782 }
5783
5784 /* Outer perimeter code */
5785
5786 /*
5787 * The outer syncq uses the fields and flags in the syncq slightly
5788 * differently from the inner syncqs.
5789 * sq_count Incremented when there are pending or running
5790 * writers at the outer perimeter to prevent the set of
5791 * inner syncqs that belong to the outer perimeter from
5792 * changing.
5793 * sq_head/tail List of deferred qwriter(OUTER) operations.
5794 *
5795 * SQ_BLOCKED Set to prevent traversing of sq_next,sq_prev while
5796 * inner syncqs are added to or removed from the
5797 * outer perimeter.
5798 * SQ_QUEUED sq_head/tail has messages or events queued.
5799 *
5800 * SQ_WRITER A thread is currently traversing all the inner syncqs
5801 * setting the SQ_WRITER flag.
5802 */
5803
5804 /*
5805 * Get write access at the outer perimeter.
5806 * Note that read access is done by entersq, putnext, and put by simply
5807 * incrementing sq_count in the inner syncq.
5808 *
5809 * Waits until "flags" is no longer set in the outer to prevent multiple
5810 * threads from having write access at the same time. SQ_WRITER has to be part
5811 * of "flags".
5812 *
5813 * Increases sq_count on the outer syncq to keep away outer_insert/remove
5814 * until the outer_exit is finished.
5815 *
5816 * outer_enter is vulnerable to starvation since it does not prevent new
5817 * threads from entering the inner syncqs while it is waiting for sq_count to
5818 * go to zero.
5819 */
5820 void
5821 outer_enter(syncq_t *outer, uint16_t flags)
5822 {
5823 syncq_t *sq;
5824 int wait_needed;
5825 uint16_t count;
5826
5827 ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5828 outer->sq_oprev != NULL);
5829 ASSERT(flags & SQ_WRITER);
5830
5831 retry:
5832 mutex_enter(SQLOCK(outer));
5833 while (outer->sq_flags & flags) {
5834 outer->sq_flags |= SQ_WANTWAKEUP;
5835 cv_wait(&outer->sq_wait, SQLOCK(outer));
5836 }
5837
5838 ASSERT(!(outer->sq_flags & SQ_WRITER));
5839 outer->sq_flags |= SQ_WRITER;
5840 outer->sq_count++;
5841 ASSERT(outer->sq_count != 0); /* wraparound */
5842 wait_needed = 0;
5843 /*
5844 * Set SQ_WRITER on all the inner syncqs while holding
5845 * the SQLOCK on the outer syncq. This ensures that the changing
5846 * of SQ_WRITER is atomic under the outer SQLOCK.
5847 */
5848 for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext) {
5849 mutex_enter(SQLOCK(sq));
5850 count = sq->sq_count;
5851 SQ_PUTLOCKS_ENTER(sq);
5852 sq->sq_flags |= SQ_WRITER;
5853 SUM_SQ_PUTCOUNTS(sq, count);
5854 if (count != 0)
5855 wait_needed = 1;
5856 SQ_PUTLOCKS_EXIT(sq);
5857 mutex_exit(SQLOCK(sq));
5858 }
5859 mutex_exit(SQLOCK(outer));
5860
5861 /*
5862 * Get everybody out of the syncqs sequentially.
5863 * Note that we don't actually need to acquire the PUTLOCKS, since
5864 * we have already cleared the fastbit, and set QWRITER. By
5865 * definition, the count can not increase since putnext will
5866 * take the slowlock path (and the purpose of acquiring the
5867 * putlocks was to make sure it didn't increase while we were
5868 * waiting).
5869 *
5870 * Note that we still acquire the PUTLOCKS to be safe.
5871 */
5872 if (wait_needed) {
5873 for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext) {
5874 mutex_enter(SQLOCK(sq));
5875 count = sq->sq_count;
5876 SQ_PUTLOCKS_ENTER(sq);
5877 SUM_SQ_PUTCOUNTS(sq, count);
5878 while (count != 0) {
5879 sq->sq_flags |= SQ_WANTWAKEUP;
5880 SQ_PUTLOCKS_EXIT(sq);
5881 cv_wait(&sq->sq_wait, SQLOCK(sq));
5882 count = sq->sq_count;
5883 SQ_PUTLOCKS_ENTER(sq);
5884 SUM_SQ_PUTCOUNTS(sq, count);
5885 }
5886 SQ_PUTLOCKS_EXIT(sq);
5887 mutex_exit(SQLOCK(sq));
5888 }
5889 /*
5890 * Verify that none of the flags got set while we
5891 * were waiting for the sq_counts to drop.
5892 * If this happens we exit and retry entering the
5893 * outer perimeter.
5894 */
5895 mutex_enter(SQLOCK(outer));
5896 if (outer->sq_flags & (flags & ~SQ_WRITER)) {
5897 mutex_exit(SQLOCK(outer));
5898 outer_exit(outer);
5899 goto retry;
5900 }
5901 mutex_exit(SQLOCK(outer));
5902 }
5903 }
5904
5905 /*
5906 * Drop the write access at the outer perimeter.
5907 * Read access is dropped implicitly (by putnext, put, and leavesq) by
5908 * decrementing sq_count.
5909 */
5910 void
5911 outer_exit(syncq_t *outer)
5912 {
5913 syncq_t *sq;
5914 int drain_needed;
5915 uint16_t flags;
5916
5917 ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5918 outer->sq_oprev != NULL);
5919 ASSERT(MUTEX_NOT_HELD(SQLOCK(outer)));
5920
5921 /*
5922 * Atomically (from the perspective of threads calling become_writer)
5923 * drop the write access at the outer perimeter by holding
5924 * SQLOCK(outer) across all the dropsq calls and the resetting of
5925 * SQ_WRITER.
5926 * This defines a locking order between the outer perimeter
5927 * SQLOCK and the inner perimeter SQLOCKs.
5928 */
5929 mutex_enter(SQLOCK(outer));
5930 flags = outer->sq_flags;
5931 ASSERT(outer->sq_flags & SQ_WRITER);
5932 if (flags & SQ_QUEUED) {
5933 write_now(outer);
5934 flags = outer->sq_flags;
5935 }
5936
5937 /*
5938 * sq_onext is stable since sq_count has not yet been decreased.
5939 * Reset the SQ_WRITER flags in all syncqs.
5940 * After dropping SQ_WRITER on the outer syncq we empty all the
5941 * inner syncqs.
5942 */
5943 drain_needed = 0;
5944 for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext)
5945 drain_needed += dropsq(sq, SQ_WRITER);
5946 ASSERT(!(outer->sq_flags & SQ_QUEUED));
5947 flags &= ~SQ_WRITER;
5948 if (drain_needed) {
5949 outer->sq_flags = flags;
5950 mutex_exit(SQLOCK(outer));
5951 for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext)
5952 emptysq(sq);
5953 mutex_enter(SQLOCK(outer));
5954 flags = outer->sq_flags;
5955 }
5956 if (flags & SQ_WANTWAKEUP) {
5957 flags &= ~SQ_WANTWAKEUP;
5958 cv_broadcast(&outer->sq_wait);
5959 }
5960 outer->sq_flags = flags;
5961 ASSERT(outer->sq_count > 0);
5962 outer->sq_count--;
5963 mutex_exit(SQLOCK(outer));
5964 }
5965
5966 /*
5967 * Add another syncq to an outer perimeter.
5968 * Block out all other access to the outer perimeter while it is being
5969 * changed using blocksq.
5970 * Assumes that the caller has *not* done an outer_enter.
5971 *
5972 * Vulnerable to starvation in blocksq.
5973 */
5974 static void
5975 outer_insert(syncq_t *outer, syncq_t *sq)
5976 {
5977 ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5978 outer->sq_oprev != NULL);
5979 ASSERT(sq->sq_outer == NULL && sq->sq_onext == NULL &&
5980 sq->sq_oprev == NULL); /* Can't be in an outer perimeter */
5981
5982 /* Get exclusive access to the outer perimeter list */
5983 blocksq(outer, SQ_BLOCKED, 0);
5984 ASSERT(outer->sq_flags & SQ_BLOCKED);
5985 ASSERT(!(outer->sq_flags & SQ_WRITER));
5986
5987 mutex_enter(SQLOCK(sq));
5988 sq->sq_outer = outer;
5989 outer->sq_onext->sq_oprev = sq;
5990 sq->sq_onext = outer->sq_onext;
5991 outer->sq_onext = sq;
5992 sq->sq_oprev = outer;
5993 mutex_exit(SQLOCK(sq));
5994 unblocksq(outer, SQ_BLOCKED, 1);
5995 }
5996
5997 /*
5998 * Remove a syncq from an outer perimeter.
5999 * Block out all other access to the outer perimeter while it is being
6000 * changed using blocksq.
6001 * Assumes that the caller has *not* done an outer_enter.
6002 *
6003 * Vulnerable to starvation in blocksq.
6004 */
6005 static void
6006 outer_remove(syncq_t *outer, syncq_t *sq)
6007 {
6008 ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
6009 outer->sq_oprev != NULL);
6010 ASSERT(sq->sq_outer == outer);
6011
6012 /* Get exclusive access to the outer perimeter list */
6013 blocksq(outer, SQ_BLOCKED, 0);
6014 ASSERT(outer->sq_flags & SQ_BLOCKED);
6015 ASSERT(!(outer->sq_flags & SQ_WRITER));
6016
6017 mutex_enter(SQLOCK(sq));
6018 sq->sq_outer = NULL;
6019 sq->sq_onext->sq_oprev = sq->sq_oprev;
6020 sq->sq_oprev->sq_onext = sq->sq_onext;
6021 sq->sq_oprev = sq->sq_onext = NULL;
6022 mutex_exit(SQLOCK(sq));
6023 unblocksq(outer, SQ_BLOCKED, 1);
6024 }
6025
6026 /*
6027 * Queue a deferred qwriter(OUTER) callback for this outer perimeter.
6028 * If this is the first callback for this outer perimeter then add
6029 * this outer perimeter to the list of outer perimeters that
6030 * the qwriter_outer_thread will process.
6031 *
6032 * Increments sq_count in the outer syncq to prevent the membership
6033 * of the outer perimeter (in terms of inner syncqs) to change while
6034 * the callback is pending.
6035 */
6036 static void
6037 queue_writer(syncq_t *outer, void (*func)(), queue_t *q, mblk_t *mp)
6038 {
6039 ASSERT(MUTEX_HELD(SQLOCK(outer)));
6040
6041 mp->b_prev = (mblk_t *)func;
6042 mp->b_queue = q;
6043 mp->b_next = NULL;
6044 outer->sq_count++; /* Decremented when dequeued */
6045 ASSERT(outer->sq_count != 0); /* Wraparound */
6046 if (outer->sq_evhead == NULL) {
6047 /* First message. */
6048 outer->sq_evhead = outer->sq_evtail = mp;
6049 outer->sq_flags |= SQ_EVENTS;
6050 mutex_exit(SQLOCK(outer));
6051 STRSTAT(qwr_outer);
6052 (void) taskq_dispatch(streams_taskq,
6053 (task_func_t *)qwriter_outer_service, outer, TQ_SLEEP);
6054 } else {
6055 ASSERT(outer->sq_flags & SQ_EVENTS);
6056 outer->sq_evtail->b_next = mp;
6057 outer->sq_evtail = mp;
6058 mutex_exit(SQLOCK(outer));
6059 }
6060 }
6061
6062 /*
6063 * Try and upgrade to write access at the outer perimeter. If this can
6064 * not be done without blocking then queue the callback to be done
6065 * by the qwriter_outer_thread.
6066 *
6067 * This routine can only be called from put or service procedures plus
6068 * asynchronous callback routines that have properly entered the queue (with
6069 * entersq). Thus qwriter(OUTER) assumes the caller has one claim on the syncq
6070 * associated with q.
6071 */
6072 void
6073 qwriter_outer(queue_t *q, mblk_t *mp, void (*func)())
6074 {
6075 syncq_t *osq, *sq, *outer;
6076 int failed;
6077 uint16_t flags;
6078
6079 osq = q->q_syncq;
6080 outer = osq->sq_outer;
6081 if (outer == NULL)
6082 panic("qwriter(PERIM_OUTER): no outer perimeter");
6083 ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
6084 outer->sq_oprev != NULL);
6085
6086 mutex_enter(SQLOCK(outer));
6087 flags = outer->sq_flags;
6088 /*
6089 * If some thread is traversing sq_next, or if we are blocked by
6090 * outer_insert or outer_remove, or if the we already have queued
6091 * callbacks, then queue this callback for later processing.
6092 *
6093 * Also queue the qwriter for an interrupt thread in order
6094 * to reduce the time spent running at high IPL.
6095 * to identify there are events.
6096 */
6097 if ((flags & SQ_GOAWAY) || (curthread->t_pri >= kpreemptpri)) {
6098 /*
6099 * Queue the become_writer request.
6100 * The queueing is atomic under SQLOCK(outer) in order
6101 * to synchronize with outer_exit.
6102 * queue_writer will drop the outer SQLOCK
6103 */
6104 if (flags & SQ_BLOCKED) {
6105 /* Must set SQ_WRITER on inner perimeter */
6106 mutex_enter(SQLOCK(osq));
6107 osq->sq_flags |= SQ_WRITER;
6108 mutex_exit(SQLOCK(osq));
6109 } else {
6110 if (!(flags & SQ_WRITER)) {
6111 /*
6112 * The outer could have been SQ_BLOCKED thus
6113 * SQ_WRITER might not be set on the inner.
6114 */
6115 mutex_enter(SQLOCK(osq));
6116 osq->sq_flags |= SQ_WRITER;
6117 mutex_exit(SQLOCK(osq));
6118 }
6119 ASSERT(osq->sq_flags & SQ_WRITER);
6120 }
6121 queue_writer(outer, func, q, mp);
6122 return;
6123 }
6124 /*
6125 * We are half-way to exclusive access to the outer perimeter.
6126 * Prevent any outer_enter, qwriter(OUTER), or outer_insert/remove
6127 * while the inner syncqs are traversed.
6128 */
6129 outer->sq_count++;
6130 ASSERT(outer->sq_count != 0); /* wraparound */
6131 flags |= SQ_WRITER;
6132 /*
6133 * Check if we can run the function immediately. Mark all
6134 * syncqs with the writer flag to prevent new entries into
6135 * put and service procedures.
6136 *
6137 * Set SQ_WRITER on all the inner syncqs while holding
6138 * the SQLOCK on the outer syncq. This ensures that the changing
6139 * of SQ_WRITER is atomic under the outer SQLOCK.
6140 */
6141 failed = 0;
6142 for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext) {
6143 uint16_t count;
6144 uint_t maxcnt = (sq == osq) ? 1 : 0;
6145
6146 mutex_enter(SQLOCK(sq));
6147 count = sq->sq_count;
6148 SQ_PUTLOCKS_ENTER(sq);
6149 SUM_SQ_PUTCOUNTS(sq, count);
6150 if (sq->sq_count > maxcnt)
6151 failed = 1;
6152 sq->sq_flags |= SQ_WRITER;
6153 SQ_PUTLOCKS_EXIT(sq);
6154 mutex_exit(SQLOCK(sq));
6155 }
6156 if (failed) {
6157 /*
6158 * Some other thread has a read claim on the outer perimeter.
6159 * Queue the callback for deferred processing.
6160 *
6161 * queue_writer will set SQ_QUEUED before we drop SQ_WRITER
6162 * so that other qwriter(OUTER) calls will queue their
6163 * callbacks as well. queue_writer increments sq_count so we
6164 * decrement to compensate for the our increment.
6165 *
6166 * Dropping SQ_WRITER enables the writer thread to work
6167 * on this outer perimeter.
6168 */
6169 outer->sq_flags = flags;
6170 queue_writer(outer, func, q, mp);
6171 /* queue_writer dropper the lock */
6172 mutex_enter(SQLOCK(outer));
6173 ASSERT(outer->sq_count > 0);
6174 outer->sq_count--;
6175 ASSERT(outer->sq_flags & SQ_WRITER);
6176 flags = outer->sq_flags;
6177 flags &= ~SQ_WRITER;
6178 if (flags & SQ_WANTWAKEUP) {
6179 flags &= ~SQ_WANTWAKEUP;
6180 cv_broadcast(&outer->sq_wait);
6181 }
6182 outer->sq_flags = flags;
6183 mutex_exit(SQLOCK(outer));
6184 return;
6185 } else {
6186 outer->sq_flags = flags;
6187 mutex_exit(SQLOCK(outer));
6188 }
6189
6190 /* Can run it immediately */
6191 (*func)(q, mp);
6192
6193 outer_exit(outer);
6194 }
6195
6196 /*
6197 * Dequeue all writer callbacks from the outer perimeter and run them.
6198 */
6199 static void
6200 write_now(syncq_t *outer)
6201 {
6202 mblk_t *mp;
6203 queue_t *q;
6204 void (*func)();
6205
6206 ASSERT(MUTEX_HELD(SQLOCK(outer)));
6207 ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
6208 outer->sq_oprev != NULL);
6209 while ((mp = outer->sq_evhead) != NULL) {
6210 /*
6211 * queues cannot be placed on the queuelist on the outer
6212 * perimeter.
6213 */
6214 ASSERT(!(outer->sq_flags & SQ_MESSAGES));
6215 ASSERT((outer->sq_flags & SQ_EVENTS));
6216
6217 outer->sq_evhead = mp->b_next;
6218 if (outer->sq_evhead == NULL) {
6219 outer->sq_evtail = NULL;
6220 outer->sq_flags &= ~SQ_EVENTS;
6221 }
6222 ASSERT(outer->sq_count != 0);
6223 outer->sq_count--; /* Incremented when enqueued. */
6224 mutex_exit(SQLOCK(outer));
6225 /*
6226 * Drop the message if the queue is closing.
6227 * Make sure that the queue is "claimed" when the callback
6228 * is run in order to satisfy various ASSERTs.
6229 */
6230 q = mp->b_queue;
6231 func = (void (*)())mp->b_prev;
6232 ASSERT(func != NULL);
6233 mp->b_next = mp->b_prev = NULL;
6234 if (q->q_flag & QWCLOSE) {
6235 freemsg(mp);
6236 } else {
6237 claimq(q);
6238 (*func)(q, mp);
6239 releaseq(q);
6240 }
6241 mutex_enter(SQLOCK(outer));
6242 }
6243 ASSERT(MUTEX_HELD(SQLOCK(outer)));
6244 }
6245
6246 /*
6247 * The list of messages on the inner syncq is effectively hashed
6248 * by destination queue. These destination queues are doubly
6249 * linked lists (hopefully) in priority order. Messages are then
6250 * put on the queue referenced by the q_sqhead/q_sqtail elements.
6251 * Additional messages are linked together by the b_next/b_prev
6252 * elements in the mblk, with (similar to putq()) the first message
6253 * having a NULL b_prev and the last message having a NULL b_next.
6254 *
6255 * Events, such as qwriter callbacks, are put onto a list in FIFO
6256 * order referenced by sq_evhead, and sq_evtail. This is a singly
6257 * linked list, and messages here MUST be processed in the order queued.
6258 */
6259
6260 /*
6261 * Run the events on the syncq event list (sq_evhead).
6262 * Assumes there is only one claim on the syncq, it is
6263 * already exclusive (SQ_EXCL set), and the SQLOCK held.
6264 * Messages here are processed in order, with the SQ_EXCL bit
6265 * held all the way through till the last message is processed.
6266 */
6267 void
6268 sq_run_events(syncq_t *sq)
6269 {
6270 mblk_t *bp;
6271 queue_t *qp;
6272 uint16_t flags = sq->sq_flags;
6273 void (*func)();
6274
6275 ASSERT(MUTEX_HELD(SQLOCK(sq)));
6276 ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6277 sq->sq_oprev == NULL) ||
6278 (sq->sq_outer != NULL && sq->sq_onext != NULL &&
6279 sq->sq_oprev != NULL));
6280
6281 ASSERT(flags & SQ_EXCL);
6282 ASSERT(sq->sq_count == 1);
6283
6284 /*
6285 * We need to process all of the events on this list. It
6286 * is possible that new events will be added while we are
6287 * away processing a callback, so on every loop, we start
6288 * back at the beginning of the list.
6289 */
6290 /*
6291 * We have to reaccess sq_evhead since there is a
6292 * possibility of a new entry while we were running
6293 * the callback.
6294 */
6295 for (bp = sq->sq_evhead; bp != NULL; bp = sq->sq_evhead) {
6296 ASSERT(bp->b_queue->q_syncq == sq);
6297 ASSERT(sq->sq_flags & SQ_EVENTS);
6298
6299 qp = bp->b_queue;
6300 func = (void (*)())bp->b_prev;
6301 ASSERT(func != NULL);
6302
6303 /*
6304 * Messages from the event queue must be taken off in
6305 * FIFO order.
6306 */
6307 ASSERT(sq->sq_evhead == bp);
6308 sq->sq_evhead = bp->b_next;
6309
6310 if (bp->b_next == NULL) {
6311 /* Deleting last */
6312 ASSERT(sq->sq_evtail == bp);
6313 sq->sq_evtail = NULL;
6314 sq->sq_flags &= ~SQ_EVENTS;
6315 }
6316 bp->b_prev = bp->b_next = NULL;
6317 ASSERT(bp->b_datap->db_ref != 0);
6318
6319 mutex_exit(SQLOCK(sq));
6320
6321 (*func)(qp, bp);
6322
6323 mutex_enter(SQLOCK(sq));
6324 /*
6325 * re-read the flags, since they could have changed.
6326 */
6327 flags = sq->sq_flags;
6328 ASSERT(flags & SQ_EXCL);
6329 }
6330 ASSERT(sq->sq_evhead == NULL && sq->sq_evtail == NULL);
6331 ASSERT(!(sq->sq_flags & SQ_EVENTS));
6332
6333 if (flags & SQ_WANTWAKEUP) {
6334 flags &= ~SQ_WANTWAKEUP;
6335 cv_broadcast(&sq->sq_wait);
6336 }
6337 if (flags & SQ_WANTEXWAKEUP) {
6338 flags &= ~SQ_WANTEXWAKEUP;
6339 cv_broadcast(&sq->sq_exitwait);
6340 }
6341 sq->sq_flags = flags;
6342 }
6343
6344 /*
6345 * Put messages on the event list.
6346 * If we can go exclusive now, do so and process the event list, otherwise
6347 * let the last claim service this list (or wake the sqthread).
6348 * This procedure assumes SQLOCK is held. To run the event list, it
6349 * must be called with no claims.
6350 */
6351 static void
6352 sqfill_events(syncq_t *sq, queue_t *q, mblk_t *mp, void (*func)())
6353 {
6354 uint16_t count;
6355
6356 ASSERT(MUTEX_HELD(SQLOCK(sq)));
6357 ASSERT(func != NULL);
6358
6359 /*
6360 * This is a callback. Add it to the list of callbacks
6361 * and see about upgrading.
6362 */
6363 mp->b_prev = (mblk_t *)func;
6364 mp->b_queue = q;
6365 mp->b_next = NULL;
6366 if (sq->sq_evhead == NULL) {
6367 sq->sq_evhead = sq->sq_evtail = mp;
6368 sq->sq_flags |= SQ_EVENTS;
6369 } else {
6370 ASSERT(sq->sq_evtail != NULL);
6371 ASSERT(sq->sq_evtail->b_next == NULL);
6372 ASSERT(sq->sq_flags & SQ_EVENTS);
6373 sq->sq_evtail->b_next = mp;
6374 sq->sq_evtail = mp;
6375 }
6376 /*
6377 * We have set SQ_EVENTS, so threads will have to
6378 * unwind out of the perimeter, and new entries will
6379 * not grab a putlock. But we still need to know
6380 * how many threads have already made a claim to the
6381 * syncq, so grab the putlocks, and sum the counts.
6382 * If there are no claims on the syncq, we can upgrade
6383 * to exclusive, and run the event list.
6384 * NOTE: We hold the SQLOCK, so we can just grab the
6385 * putlocks.
6386 */
6387 count = sq->sq_count;
6388 SQ_PUTLOCKS_ENTER(sq);
6389 SUM_SQ_PUTCOUNTS(sq, count);
6390 /*
6391 * We have no claim, so we need to check if there
6392 * are no others, then we can upgrade.
6393 */
6394 /*
6395 * There are currently no claims on
6396 * the syncq by this thread (at least on this entry). The thread who has
6397 * the claim should drain syncq.
6398 */
6399 if (count > 0) {
6400 /*
6401 * Can't upgrade - other threads inside.
6402 */
6403 SQ_PUTLOCKS_EXIT(sq);
6404 mutex_exit(SQLOCK(sq));
6405 return;
6406 }
6407 /*
6408 * Need to set SQ_EXCL and make a claim on the syncq.
6409 */
6410 ASSERT((sq->sq_flags & SQ_EXCL) == 0);
6411 sq->sq_flags |= SQ_EXCL;
6412 ASSERT(sq->sq_count == 0);
6413 sq->sq_count++;
6414 SQ_PUTLOCKS_EXIT(sq);
6415
6416 /* Process the events list */
6417 sq_run_events(sq);
6418
6419 /*
6420 * Release our claim...
6421 */
6422 sq->sq_count--;
6423
6424 /*
6425 * And release SQ_EXCL.
6426 * We don't need to acquire the putlocks to release
6427 * SQ_EXCL, since we are exclusive, and hold the SQLOCK.
6428 */
6429 sq->sq_flags &= ~SQ_EXCL;
6430
6431 /*
6432 * sq_run_events should have released SQ_EXCL
6433 */
6434 ASSERT(!(sq->sq_flags & SQ_EXCL));
6435
6436 /*
6437 * If anything happened while we were running the
6438 * events (or was there before), we need to process
6439 * them now. We shouldn't be exclusive sine we
6440 * released the perimeter above (plus, we asserted
6441 * for it).
6442 */
6443 if (!(sq->sq_flags & SQ_STAYAWAY) && (sq->sq_flags & SQ_QUEUED))
6444 drain_syncq(sq);
6445 else
6446 mutex_exit(SQLOCK(sq));
6447 }
6448
6449 /*
6450 * Perform delayed processing. The caller has to make sure that it is safe
6451 * to enter the syncq (e.g. by checking that none of the SQ_STAYAWAY bits are
6452 * set).
6453 *
6454 * Assume that the caller has NO claims on the syncq. However, a claim
6455 * on the syncq does not indicate that a thread is draining the syncq.
6456 * There may be more claims on the syncq than there are threads draining
6457 * (i.e. #_threads_draining <= sq_count)
6458 *
6459 * drain_syncq has to terminate when one of the SQ_STAYAWAY bits gets set
6460 * in order to preserve qwriter(OUTER) ordering constraints.
6461 *
6462 * sq_putcount only needs to be checked when dispatching the queued
6463 * writer call for CIPUT sync queue, but this is handled in sq_run_events.
6464 */
6465 void
6466 drain_syncq(syncq_t *sq)
6467 {
6468 queue_t *qp;
6469 uint16_t count;
6470 uint16_t type = sq->sq_type;
6471 uint16_t flags = sq->sq_flags;
6472 boolean_t bg_service = sq->sq_svcflags & SQ_SERVICE;
6473
6474 TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_START,
6475 "drain_syncq start:%p", sq);
6476 ASSERT(MUTEX_HELD(SQLOCK(sq)));
6477 ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6478 sq->sq_oprev == NULL) ||
6479 (sq->sq_outer != NULL && sq->sq_onext != NULL &&
6480 sq->sq_oprev != NULL));
6481
6482 /*
6483 * Drop SQ_SERVICE flag.
6484 */
6485 if (bg_service)
6486 sq->sq_svcflags &= ~SQ_SERVICE;
6487
6488 /*
6489 * If SQ_EXCL is set, someone else is processing this syncq - let him
6490 * finish the job.
6491 */
6492 if (flags & SQ_EXCL) {
6493 if (bg_service) {
6494 ASSERT(sq->sq_servcount != 0);
6495 sq->sq_servcount--;
6496 }
6497 mutex_exit(SQLOCK(sq));
6498 return;
6499 }
6500
6501 /*
6502 * This routine can be called by a background thread if
6503 * it was scheduled by a hi-priority thread. SO, if there are
6504 * NOT messages queued, return (remember, we have the SQLOCK,
6505 * and it cannot change until we release it). Wakeup any waiters also.
6506 */
6507 if (!(flags & SQ_QUEUED)) {
6508 if (flags & SQ_WANTWAKEUP) {
6509 flags &= ~SQ_WANTWAKEUP;
6510 cv_broadcast(&sq->sq_wait);
6511 }
6512 if (flags & SQ_WANTEXWAKEUP) {
6513 flags &= ~SQ_WANTEXWAKEUP;
6514 cv_broadcast(&sq->sq_exitwait);
6515 }
6516 sq->sq_flags = flags;
6517 if (bg_service) {
6518 ASSERT(sq->sq_servcount != 0);
6519 sq->sq_servcount--;
6520 }
6521 mutex_exit(SQLOCK(sq));
6522 return;
6523 }
6524
6525 /*
6526 * If this is not a concurrent put perimeter, we need to
6527 * become exclusive to drain. Also, if not CIPUT, we would
6528 * not have acquired a putlock, so we don't need to check
6529 * the putcounts. If not entering with a claim, we test
6530 * for sq_count == 0.
6531 */
6532 type = sq->sq_type;
6533 if (!(type & SQ_CIPUT)) {
6534 if (sq->sq_count > 1) {
6535 if (bg_service) {
6536 ASSERT(sq->sq_servcount != 0);
6537 sq->sq_servcount--;
6538 }
6539 mutex_exit(SQLOCK(sq));
6540 return;
6541 }
6542 sq->sq_flags |= SQ_EXCL;
6543 }
6544
6545 /*
6546 * This is where we make a claim to the syncq.
6547 * This can either be done by incrementing a putlock, or
6548 * the sq_count. But since we already have the SQLOCK
6549 * here, we just bump the sq_count.
6550 *
6551 * Note that after we make a claim, we need to let the code
6552 * fall through to the end of this routine to clean itself
6553 * up. A return in the while loop will put the syncq in a
6554 * very bad state.
6555 */
6556 sq->sq_count++;
6557 ASSERT(sq->sq_count != 0); /* wraparound */
6558
6559 while ((flags = sq->sq_flags) & SQ_QUEUED) {
6560 /*
6561 * If we are told to stayaway or went exclusive,
6562 * we are done.
6563 */
6564 if (flags & (SQ_STAYAWAY)) {
6565 break;
6566 }
6567
6568 /*
6569 * If there are events to run, do so.
6570 * We have one claim to the syncq, so if there are
6571 * more than one, other threads are running.
6572 */
6573 if (sq->sq_evhead != NULL) {
6574 ASSERT(sq->sq_flags & SQ_EVENTS);
6575
6576 count = sq->sq_count;
6577 SQ_PUTLOCKS_ENTER(sq);
6578 SUM_SQ_PUTCOUNTS(sq, count);
6579 if (count > 1) {
6580 SQ_PUTLOCKS_EXIT(sq);
6581 /* Can't upgrade - other threads inside */
6582 break;
6583 }
6584 ASSERT((flags & SQ_EXCL) == 0);
6585 sq->sq_flags = flags | SQ_EXCL;
6586 SQ_PUTLOCKS_EXIT(sq);
6587 /*
6588 * we have the only claim, run the events,
6589 * sq_run_events will clear the SQ_EXCL flag.
6590 */
6591 sq_run_events(sq);
6592
6593 /*
6594 * If this is a CIPUT perimeter, we need
6595 * to drop the SQ_EXCL flag so we can properly
6596 * continue draining the syncq.
6597 */
6598 if (type & SQ_CIPUT) {
6599 ASSERT(sq->sq_flags & SQ_EXCL);
6600 sq->sq_flags &= ~SQ_EXCL;
6601 }
6602
6603 /*
6604 * And go back to the beginning just in case
6605 * anything changed while we were away.
6606 */
6607 ASSERT((sq->sq_flags & SQ_EXCL) || (type & SQ_CIPUT));
6608 continue;
6609 }
6610
6611 ASSERT(sq->sq_evhead == NULL);
6612 ASSERT(!(sq->sq_flags & SQ_EVENTS));
6613
6614 /*
6615 * Find the queue that is not draining.
6616 *
6617 * q_draining is protected by QLOCK which we do not hold.
6618 * But if it was set, then a thread was draining, and if it gets
6619 * cleared, then it was because the thread has successfully
6620 * drained the syncq, or a GOAWAY state occurred. For the GOAWAY
6621 * state to happen, a thread needs the SQLOCK which we hold, and
6622 * if there was such a flag, we would have already seen it.
6623 */
6624
6625 for (qp = sq->sq_head;
6626 qp != NULL && (qp->q_draining ||
6627 (qp->q_sqflags & Q_SQDRAINING));
6628 qp = qp->q_sqnext)
6629 ;
6630
6631 if (qp == NULL)
6632 break;
6633
6634 /*
6635 * We have a queue to work on, and we hold the
6636 * SQLOCK and one claim, call qdrain_syncq.
6637 * This means we need to release the SQLOCK and
6638 * acquire the QLOCK (OK since we have a claim).
6639 * Note that qdrain_syncq will actually dequeue
6640 * this queue from the sq_head list when it is
6641 * convinced all the work is done and release
6642 * the QLOCK before returning.
6643 */
6644 qp->q_sqflags |= Q_SQDRAINING;
6645 mutex_exit(SQLOCK(sq));
6646 mutex_enter(QLOCK(qp));
6647 qdrain_syncq(sq, qp);
6648 mutex_enter(SQLOCK(sq));
6649
6650 /* The queue is drained */
6651 ASSERT(qp->q_sqflags & Q_SQDRAINING);
6652 qp->q_sqflags &= ~Q_SQDRAINING;
6653 /*
6654 * NOTE: After this point qp should not be used since it may be
6655 * closed.
6656 */
6657 }
6658
6659 ASSERT(MUTEX_HELD(SQLOCK(sq)));
6660 flags = sq->sq_flags;
6661
6662 /*
6663 * sq->sq_head cannot change because we hold the
6664 * sqlock. However, a thread CAN decide that it is no longer
6665 * going to drain that queue. However, this should be due to
6666 * a GOAWAY state, and we should see that here.
6667 *
6668 * This loop is not very efficient. One solution may be adding a second
6669 * pointer to the "draining" queue, but it is difficult to do when
6670 * queues are inserted in the middle due to priority ordering. Another
6671 * possibility is to yank the queue out of the sq list and put it onto
6672 * the "draining list" and then put it back if it can't be drained.
6673 */
6674
6675 ASSERT((sq->sq_head == NULL) || (flags & SQ_GOAWAY) ||
6676 (type & SQ_CI) || sq->sq_head->q_draining);
6677
6678 /* Drop SQ_EXCL for non-CIPUT perimeters */
6679 if (!(type & SQ_CIPUT))
6680 flags &= ~SQ_EXCL;
6681 ASSERT((flags & SQ_EXCL) == 0);
6682
6683 /* Wake up any waiters. */
6684 if (flags & SQ_WANTWAKEUP) {
6685 flags &= ~SQ_WANTWAKEUP;
6686 cv_broadcast(&sq->sq_wait);
6687 }
6688 if (flags & SQ_WANTEXWAKEUP) {
6689 flags &= ~SQ_WANTEXWAKEUP;
6690 cv_broadcast(&sq->sq_exitwait);
6691 }
6692 sq->sq_flags = flags;
6693
6694 ASSERT(sq->sq_count != 0);
6695 /* Release our claim. */
6696 sq->sq_count--;
6697
6698 if (bg_service) {
6699 ASSERT(sq->sq_servcount != 0);
6700 sq->sq_servcount--;
6701 }
6702
6703 mutex_exit(SQLOCK(sq));
6704
6705 TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_END,
6706 "drain_syncq end:%p", sq);
6707 }
6708
6709
6710 /*
6711 *
6712 * qdrain_syncq can be called (currently) from only one of two places:
6713 * drain_syncq
6714 * putnext (or some variation of it).
6715 * and eventually
6716 * qwait(_sig)
6717 *
6718 * If called from drain_syncq, we found it in the list of queues needing
6719 * service, so there is work to be done (or it wouldn't be in the list).
6720 *
6721 * If called from some putnext variation, it was because the
6722 * perimeter is open, but messages are blocking a putnext and
6723 * there is not a thread working on it. Now a thread could start
6724 * working on it while we are getting ready to do so ourself, but
6725 * the thread would set the q_draining flag, and we can spin out.
6726 *
6727 * As for qwait(_sig), I think I shall let it continue to call
6728 * drain_syncq directly (after all, it will get here eventually).
6729 *
6730 * qdrain_syncq has to terminate when:
6731 * - one of the SQ_STAYAWAY bits gets set to preserve qwriter(OUTER) ordering
6732 * - SQ_EVENTS gets set to preserve qwriter(INNER) ordering
6733 *
6734 * ASSUMES:
6735 * One claim
6736 * QLOCK held
6737 * SQLOCK not held
6738 * Will release QLOCK before returning
6739 */
6740 void
6741 qdrain_syncq(syncq_t *sq, queue_t *q)
6742 {
6743 mblk_t *bp;
6744 #ifdef DEBUG
6745 uint16_t count;
6746 #endif
6747
6748 TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_START,
6749 "drain_syncq start:%p", sq);
6750 ASSERT(q->q_syncq == sq);
6751 ASSERT(MUTEX_HELD(QLOCK(q)));
6752 ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
6753 /*
6754 * For non-CIPUT perimeters, we should be called with the exclusive bit
6755 * set already. For CIPUT perimeters, we will be doing a concurrent
6756 * drain, so it better not be set.
6757 */
6758 ASSERT((sq->sq_flags & (SQ_EXCL|SQ_CIPUT)));
6759 ASSERT(!((sq->sq_type & SQ_CIPUT) && (sq->sq_flags & SQ_EXCL)));
6760 ASSERT((sq->sq_type & SQ_CIPUT) || (sq->sq_flags & SQ_EXCL));
6761 /*
6762 * All outer pointers are set, or none of them are
6763 */
6764 ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6765 sq->sq_oprev == NULL) ||
6766 (sq->sq_outer != NULL && sq->sq_onext != NULL &&
6767 sq->sq_oprev != NULL));
6768 #ifdef DEBUG
6769 count = sq->sq_count;
6770 /*
6771 * This is OK without the putlocks, because we have one
6772 * claim either from the sq_count, or a putcount. We could
6773 * get an erroneous value from other counts, but ours won't
6774 * change, so one way or another, we will have at least a
6775 * value of one.
6776 */
6777 SUM_SQ_PUTCOUNTS(sq, count);
6778 ASSERT(count >= 1);
6779 #endif /* DEBUG */
6780
6781 /*
6782 * The first thing to do is find out if a thread is already draining
6783 * this queue. If so, we are done, just return.
6784 */
6785 if (q->q_draining) {
6786 mutex_exit(QLOCK(q));
6787 return;
6788 }
6789
6790 /*
6791 * If the perimeter is exclusive, there is nothing we can do right now,
6792 * go away. Note that there is nothing to prevent this case from
6793 * changing right after this check, but the spin-out will catch it.
6794 */
6795
6796 /* Tell other threads that we are draining this queue */
6797 q->q_draining = 1; /* Protected by QLOCK */
6798
6799 /*
6800 * If there is nothing to do, clear QFULL as necessary. This caters for
6801 * the case where an empty queue was enqueued onto the syncq.
6802 */
6803 if (q->q_sqhead == NULL) {
6804 ASSERT(q->q_syncqmsgs == 0);
6805 mutex_exit(QLOCK(q));
6806 clr_qfull(q);
6807 mutex_enter(QLOCK(q));
6808 }
6809
6810 /*
6811 * Note that q_sqhead must be re-checked here in case another message
6812 * was enqueued whilst QLOCK was dropped during the call to clr_qfull.
6813 */
6814 for (bp = q->q_sqhead; bp != NULL; bp = q->q_sqhead) {
6815 /*
6816 * Because we can enter this routine just because a putnext is
6817 * blocked, we need to spin out if the perimeter wants to go
6818 * exclusive as well as just blocked. We need to spin out also
6819 * if events are queued on the syncq.
6820 * Don't check for SQ_EXCL, because non-CIPUT perimeters would
6821 * set it, and it can't become exclusive while we hold a claim.
6822 */
6823 if (sq->sq_flags & (SQ_STAYAWAY | SQ_EVENTS)) {
6824 break;
6825 }
6826
6827 #ifdef DEBUG
6828 /*
6829 * Since we are in qdrain_syncq, we already know the queue,
6830 * but for sanity, we want to check this against the qp that
6831 * was passed in by bp->b_queue.
6832 */
6833
6834 ASSERT(bp->b_queue == q);
6835 ASSERT(bp->b_queue->q_syncq == sq);
6836 bp->b_queue = NULL;
6837
6838 /*
6839 * We would have the following check in the DEBUG code:
6840 *
6841 * if (bp->b_prev != NULL) {
6842 * ASSERT(bp->b_prev == (void (*)())q->q_qinfo->qi_putp);
6843 * }
6844 *
6845 * This can't be done, however, since IP modifies qinfo
6846 * structure at run-time (switching between IPv4 qinfo and IPv6
6847 * qinfo), invalidating the check.
6848 * So the assignment to func is left here, but the ASSERT itself
6849 * is removed until the whole issue is resolved.
6850 */
6851 #endif
6852 ASSERT(q->q_sqhead == bp);
6853 q->q_sqhead = bp->b_next;
6854 bp->b_prev = bp->b_next = NULL;
6855 ASSERT(q->q_syncqmsgs > 0);
6856 mutex_exit(QLOCK(q));
6857
6858 ASSERT(bp->b_datap->db_ref != 0);
6859
6860 (void) (*q->q_qinfo->qi_putp)(q, bp);
6861
6862 mutex_enter(QLOCK(q));
6863
6864 /*
6865 * q_syncqmsgs should only be decremented after executing the
6866 * put procedure to avoid message re-ordering. This is due to an
6867 * optimisation in putnext() which can call the put procedure
6868 * directly if it sees q_syncqmsgs == 0 (despite Q_SQQUEUED
6869 * being set).
6870 *
6871 * We also need to clear QFULL in the next service procedure
6872 * queue if this is the last message destined for that queue.
6873 *
6874 * It would make better sense to have some sort of tunable for
6875 * the low water mark, but these semantics are not yet defined.
6876 * So, alas, we use a constant.
6877 */
6878 if (--q->q_syncqmsgs == 0) {
6879 mutex_exit(QLOCK(q));
6880 clr_qfull(q);
6881 mutex_enter(QLOCK(q));
6882 }
6883
6884 /*
6885 * Always clear SQ_EXCL when CIPUT in order to handle
6886 * qwriter(INNER). The putp() can call qwriter and get exclusive
6887 * access IFF this is the only claim. So, we need to test for
6888 * this possibility, acquire the mutex and clear the bit.
6889 */
6890 if ((sq->sq_type & SQ_CIPUT) && (sq->sq_flags & SQ_EXCL)) {
6891 mutex_enter(SQLOCK(sq));
6892 sq->sq_flags &= ~SQ_EXCL;
6893 mutex_exit(SQLOCK(sq));
6894 }
6895 }
6896
6897 /*
6898 * We should either have no messages on this queue, or we were told to
6899 * goaway by a waiter (which we will wake up at the end of this
6900 * function).
6901 */
6902 ASSERT((q->q_sqhead == NULL) ||
6903 (sq->sq_flags & (SQ_STAYAWAY | SQ_EVENTS)));
6904
6905 ASSERT(MUTEX_HELD(QLOCK(q)));
6906 ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
6907
6908 /* Remove the q from the syncq list if all the messages are drained. */
6909 if (q->q_sqhead == NULL) {
6910 ASSERT(q->q_syncqmsgs == 0);
6911 mutex_enter(SQLOCK(sq));
6912 if (q->q_sqflags & Q_SQQUEUED)
6913 SQRM_Q(sq, q);
6914 mutex_exit(SQLOCK(sq));
6915 /*
6916 * Since the queue is removed from the list, reset its priority.
6917 */
6918 q->q_spri = 0;
6919 }
6920
6921 /*
6922 * Remember, the q_draining flag is used to let another thread know
6923 * that there is a thread currently draining the messages for a queue.
6924 * Since we are now done with this queue (even if there may be messages
6925 * still there), we need to clear this flag so some thread will work on
6926 * it if needed.
6927 */
6928 ASSERT(q->q_draining);
6929 q->q_draining = 0;
6930
6931 /* Called with a claim, so OK to drop all locks. */
6932 mutex_exit(QLOCK(q));
6933
6934 TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_END,
6935 "drain_syncq end:%p", sq);
6936 }
6937 /* END OF QDRAIN_SYNCQ */
6938
6939
6940 /*
6941 * This is the mate to qdrain_syncq, except that it is putting the message onto
6942 * the queue instead of draining. Since the message is destined for the queue
6943 * that is selected, there is no need to identify the function because the
6944 * message is intended for the put routine for the queue. For debug kernels,
6945 * this routine will do it anyway just in case.
6946 *
6947 * After the message is enqueued on the syncq, it calls putnext_tail()
6948 * which will schedule a background thread to actually process the message.
6949 *
6950 * Assumes that there is a claim on the syncq (sq->sq_count > 0) and
6951 * SQLOCK(sq) and QLOCK(q) are not held.
6952 */
6953 void
6954 qfill_syncq(syncq_t *sq, queue_t *q, mblk_t *mp)
6955 {
6956 ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
6957 ASSERT(MUTEX_NOT_HELD(QLOCK(q)));
6958 ASSERT(sq->sq_count > 0);
6959 ASSERT(q->q_syncq == sq);
6960 ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6961 sq->sq_oprev == NULL) ||
6962 (sq->sq_outer != NULL && sq->sq_onext != NULL &&
6963 sq->sq_oprev != NULL));
6964
6965 mutex_enter(QLOCK(q));
6966
6967 #ifdef DEBUG
6968 /*
6969 * This is used for debug in the qfill_syncq/qdrain_syncq case
6970 * to trace the queue that the message is intended for. Note
6971 * that the original use was to identify the queue and function
6972 * to call on the drain. In the new syncq, we have the context
6973 * of the queue that we are draining, so call it's putproc and
6974 * don't rely on the saved values. But for debug this is still
6975 * useful information.
6976 */
6977 mp->b_prev = (mblk_t *)q->q_qinfo->qi_putp;
6978 mp->b_queue = q;
6979 mp->b_next = NULL;
6980 #endif
6981 ASSERT(q->q_syncq == sq);
6982 /*
6983 * Enqueue the message on the list.
6984 * SQPUT_MP() accesses q_syncqmsgs. We are already holding QLOCK to
6985 * protect it. So it's ok to acquire SQLOCK after SQPUT_MP().
6986 */
6987 SQPUT_MP(q, mp);
6988 mutex_enter(SQLOCK(sq));
6989
6990 /*
6991 * And queue on syncq for scheduling, if not already queued.
6992 * Note that we need the SQLOCK for this, and for testing flags
6993 * at the end to see if we will drain. So grab it now, and
6994 * release it before we call qdrain_syncq or return.
6995 */
6996 if (!(q->q_sqflags & Q_SQQUEUED)) {
6997 q->q_spri = curthread->t_pri;
6998 SQPUT_Q(sq, q);
6999 }
7000 #ifdef DEBUG
7001 else {
7002 /*
7003 * All of these conditions MUST be true!
7004 */
7005 ASSERT(sq->sq_tail != NULL);
7006 if (sq->sq_tail == sq->sq_head) {
7007 ASSERT((q->q_sqprev == NULL) &&
7008 (q->q_sqnext == NULL));
7009 } else {
7010 ASSERT((q->q_sqprev != NULL) ||
7011 (q->q_sqnext != NULL));
7012 }
7013 ASSERT(sq->sq_flags & SQ_QUEUED);
7014 ASSERT(q->q_syncqmsgs != 0);
7015 ASSERT(q->q_sqflags & Q_SQQUEUED);
7016 }
7017 #endif
7018 mutex_exit(QLOCK(q));
7019 /*
7020 * SQLOCK is still held, so sq_count can be safely decremented.
7021 */
7022 sq->sq_count--;
7023
7024 putnext_tail(sq, q, 0);
7025 /* Should not reference sq or q after this point. */
7026 }
7027
7028 /* End of qfill_syncq */
7029
7030 /*
7031 * Remove all messages from a syncq (if qp is NULL) or remove all messages
7032 * that would be put into qp by drain_syncq.
7033 * Used when deleting the syncq (qp == NULL) or when detaching
7034 * a queue (qp != NULL).
7035 * Return non-zero if one or more messages were freed.
7036 *
7037 * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
7038 * sq_putlocks are used.
7039 *
7040 * NOTE: This function assumes that it is called from the close() context and
7041 * that all the queues in the syncq are going away. For this reason it doesn't
7042 * acquire QLOCK for modifying q_sqhead/q_sqtail fields. This assumption is
7043 * currently valid, but it is useful to rethink this function to behave properly
7044 * in other cases.
7045 */
7046 int
7047 flush_syncq(syncq_t *sq, queue_t *qp)
7048 {
7049 mblk_t *bp, *mp_head, *mp_next, *mp_prev;
7050 queue_t *q;
7051 int ret = 0;
7052
7053 mutex_enter(SQLOCK(sq));
7054
7055 /*
7056 * Before we leave, we need to make sure there are no
7057 * events listed for this queue. All events for this queue
7058 * will just be freed.
7059 */
7060 if (qp != NULL && sq->sq_evhead != NULL) {
7061 ASSERT(sq->sq_flags & SQ_EVENTS);
7062
7063 mp_prev = NULL;
7064 for (bp = sq->sq_evhead; bp != NULL; bp = mp_next) {
7065 mp_next = bp->b_next;
7066 if (bp->b_queue == qp) {
7067 /* Delete this message */
7068 if (mp_prev != NULL) {
7069 mp_prev->b_next = mp_next;
7070 /*
7071 * Update sq_evtail if the last element
7072 * is removed.
7073 */
7074 if (bp == sq->sq_evtail) {
7075 ASSERT(mp_next == NULL);
7076 sq->sq_evtail = mp_prev;
7077 }
7078 } else
7079 sq->sq_evhead = mp_next;
7080 if (sq->sq_evhead == NULL)
7081 sq->sq_flags &= ~SQ_EVENTS;
7082 bp->b_prev = bp->b_next = NULL;
7083 freemsg(bp);
7084 ret++;
7085 } else {
7086 mp_prev = bp;
7087 }
7088 }
7089 }
7090
7091 /*
7092 * Walk sq_head and:
7093 * - match qp if qp is set, remove it's messages
7094 * - all if qp is not set
7095 */
7096 q = sq->sq_head;
7097 while (q != NULL) {
7098 ASSERT(q->q_syncq == sq);
7099 if ((qp == NULL) || (qp == q)) {
7100 /*
7101 * Yank the messages as a list off the queue
7102 */
7103 mp_head = q->q_sqhead;
7104 /*
7105 * We do not have QLOCK(q) here (which is safe due to
7106 * assumptions mentioned above). To obtain the lock we
7107 * need to release SQLOCK which may allow lots of things
7108 * to change upon us. This place requires more analysis.
7109 */
7110 q->q_sqhead = q->q_sqtail = NULL;
7111 ASSERT(mp_head->b_queue &&
7112 mp_head->b_queue->q_syncq == sq);
7113
7114 /*
7115 * Free each of the messages.
7116 */
7117 for (bp = mp_head; bp != NULL; bp = mp_next) {
7118 mp_next = bp->b_next;
7119 bp->b_prev = bp->b_next = NULL;
7120 freemsg(bp);
7121 ret++;
7122 }
7123 /*
7124 * Now remove the queue from the syncq.
7125 */
7126 ASSERT(q->q_sqflags & Q_SQQUEUED);
7127 SQRM_Q(sq, q);
7128 q->q_spri = 0;
7129 q->q_syncqmsgs = 0;
7130
7131 /*
7132 * If qp was specified, we are done with it and are
7133 * going to drop SQLOCK(sq) and return. We wakeup syncq
7134 * waiters while we still have the SQLOCK.
7135 */
7136 if ((qp != NULL) && (sq->sq_flags & SQ_WANTWAKEUP)) {
7137 sq->sq_flags &= ~SQ_WANTWAKEUP;
7138 cv_broadcast(&sq->sq_wait);
7139 }
7140 /* Drop SQLOCK across clr_qfull */
7141 mutex_exit(SQLOCK(sq));
7142
7143 /*
7144 * We avoid doing the test that drain_syncq does and
7145 * unconditionally clear qfull for every flushed
7146 * message. Since flush_syncq is only called during
7147 * close this should not be a problem.
7148 */
7149 clr_qfull(q);
7150 if (qp != NULL) {
7151 return (ret);
7152 } else {
7153 mutex_enter(SQLOCK(sq));
7154 /*
7155 * The head was removed by SQRM_Q above.
7156 * reread the new head and flush it.
7157 */
7158 q = sq->sq_head;
7159 }
7160 } else {
7161 q = q->q_sqnext;
7162 }
7163 ASSERT(MUTEX_HELD(SQLOCK(sq)));
7164 }
7165
7166 if (sq->sq_flags & SQ_WANTWAKEUP) {
7167 sq->sq_flags &= ~SQ_WANTWAKEUP;
7168 cv_broadcast(&sq->sq_wait);
7169 }
7170
7171 mutex_exit(SQLOCK(sq));
7172 return (ret);
7173 }
7174
7175 /*
7176 * Propagate all messages from a syncq to the next syncq that are associated
7177 * with the specified queue. If the queue is attached to a driver or if the
7178 * messages have been added due to a qwriter(PERIM_INNER), free the messages.
7179 *
7180 * Assumes that the stream is strlock()'ed. We don't come here if there
7181 * are no messages to propagate.
7182 *
7183 * NOTE : If the queue is attached to a driver, all the messages are freed
7184 * as there is no point in propagating the messages from the driver syncq
7185 * to the closing stream head which will in turn get freed later.
7186 */
7187 static int
7188 propagate_syncq(queue_t *qp)
7189 {
7190 mblk_t *bp, *head, *tail, *prev, *next;
7191 syncq_t *sq;
7192 queue_t *nqp;
7193 syncq_t *nsq;
7194 boolean_t isdriver;
7195 int moved = 0;
7196 uint16_t flags;
7197 pri_t priority = curthread->t_pri;
7198 #ifdef DEBUG
7199 void (*func)();
7200 #endif
7201
7202 sq = qp->q_syncq;
7203 ASSERT(MUTEX_HELD(SQLOCK(sq)));
7204 /* debug macro */
7205 SQ_PUTLOCKS_HELD(sq);
7206 /*
7207 * As entersq() does not increment the sq_count for
7208 * the write side, check sq_count for non-QPERQ
7209 * perimeters alone.
7210 */
7211 ASSERT((qp->q_flag & QPERQ) || (sq->sq_count >= 1));
7212
7213 /*
7214 * propagate_syncq() can be called because of either messages on the
7215 * queue syncq or because on events on the queue syncq. Do actual
7216 * message propagations if there are any messages.
7217 */
7218 if (qp->q_syncqmsgs) {
7219 isdriver = (qp->q_flag & QISDRV);
7220
7221 if (!isdriver) {
7222 nqp = qp->q_next;
7223 nsq = nqp->q_syncq;
7224 ASSERT(MUTEX_HELD(SQLOCK(nsq)));
7225 /* debug macro */
7226 SQ_PUTLOCKS_HELD(nsq);
7227 #ifdef DEBUG
7228 func = (void (*)())nqp->q_qinfo->qi_putp;
7229 #endif
7230 }
7231
7232 SQRM_Q(sq, qp);
7233 priority = MAX(qp->q_spri, priority);
7234 qp->q_spri = 0;
7235 head = qp->q_sqhead;
7236 tail = qp->q_sqtail;
7237 qp->q_sqhead = qp->q_sqtail = NULL;
7238 qp->q_syncqmsgs = 0;
7239
7240 /*
7241 * Walk the list of messages, and free them if this is a driver,
7242 * otherwise reset the b_prev and b_queue value to the new putp.
7243 * Afterward, we will just add the head to the end of the next
7244 * syncq, and point the tail to the end of this one.
7245 */
7246
7247 for (bp = head; bp != NULL; bp = next) {
7248 next = bp->b_next;
7249 if (isdriver) {
7250 bp->b_prev = bp->b_next = NULL;
7251 freemsg(bp);
7252 continue;
7253 }
7254 /* Change the q values for this message */
7255 bp->b_queue = nqp;
7256 #ifdef DEBUG
7257 bp->b_prev = (mblk_t *)func;
7258 #endif
7259 moved++;
7260 }
7261 /*
7262 * Attach list of messages to the end of the new queue (if there
7263 * is a list of messages).
7264 */
7265
7266 if (!isdriver && head != NULL) {
7267 ASSERT(tail != NULL);
7268 if (nqp->q_sqhead == NULL) {
7269 nqp->q_sqhead = head;
7270 } else {
7271 ASSERT(nqp->q_sqtail != NULL);
7272 nqp->q_sqtail->b_next = head;
7273 }
7274 nqp->q_sqtail = tail;
7275 /*
7276 * When messages are moved from high priority queue to
7277 * another queue, the destination queue priority is
7278 * upgraded.
7279 */
7280
7281 if (priority > nqp->q_spri)
7282 nqp->q_spri = priority;
7283
7284 SQPUT_Q(nsq, nqp);
7285
7286 nqp->q_syncqmsgs += moved;
7287 ASSERT(nqp->q_syncqmsgs != 0);
7288 }
7289 }
7290
7291 /*
7292 * Before we leave, we need to make sure there are no
7293 * events listed for this queue. All events for this queue
7294 * will just be freed.
7295 */
7296 if (sq->sq_evhead != NULL) {
7297 ASSERT(sq->sq_flags & SQ_EVENTS);
7298 prev = NULL;
7299 for (bp = sq->sq_evhead; bp != NULL; bp = next) {
7300 next = bp->b_next;
7301 if (bp->b_queue == qp) {
7302 /* Delete this message */
7303 if (prev != NULL) {
7304 prev->b_next = next;
7305 /*
7306 * Update sq_evtail if the last element
7307 * is removed.
7308 */
7309 if (bp == sq->sq_evtail) {
7310 ASSERT(next == NULL);
7311 sq->sq_evtail = prev;
7312 }
7313 } else
7314 sq->sq_evhead = next;
7315 if (sq->sq_evhead == NULL)
7316 sq->sq_flags &= ~SQ_EVENTS;
7317 bp->b_prev = bp->b_next = NULL;
7318 freemsg(bp);
7319 } else {
7320 prev = bp;
7321 }
7322 }
7323 }
7324
7325 flags = sq->sq_flags;
7326
7327 /* Wake up any waiter before leaving. */
7328 if (flags & SQ_WANTWAKEUP) {
7329 flags &= ~SQ_WANTWAKEUP;
7330 cv_broadcast(&sq->sq_wait);
7331 }
7332 sq->sq_flags = flags;
7333
7334 return (moved);
7335 }
7336
7337 /*
7338 * Try and upgrade to exclusive access at the inner perimeter. If this can
7339 * not be done without blocking then request will be queued on the syncq
7340 * and drain_syncq will run it later.
7341 *
7342 * This routine can only be called from put or service procedures plus
7343 * asynchronous callback routines that have properly entered the queue (with
7344 * entersq). Thus qwriter_inner assumes the caller has one claim on the syncq
7345 * associated with q.
7346 */
7347 void
7348 qwriter_inner(queue_t *q, mblk_t *mp, void (*func)())
7349 {
7350 syncq_t *sq = q->q_syncq;
7351 uint16_t count;
7352
7353 mutex_enter(SQLOCK(sq));
7354 count = sq->sq_count;
7355 SQ_PUTLOCKS_ENTER(sq);
7356 SUM_SQ_PUTCOUNTS(sq, count);
7357 ASSERT(count >= 1);
7358 ASSERT(sq->sq_type & (SQ_CIPUT|SQ_CISVC));
7359
7360 if (count == 1) {
7361 /*
7362 * Can upgrade. This case also handles nested qwriter calls
7363 * (when the qwriter callback function calls qwriter). In that
7364 * case SQ_EXCL is already set.
7365 */
7366 sq->sq_flags |= SQ_EXCL;
7367 SQ_PUTLOCKS_EXIT(sq);
7368 mutex_exit(SQLOCK(sq));
7369 (*func)(q, mp);
7370 /*
7371 * Assumes that leavesq, putnext, and drain_syncq will reset
7372 * SQ_EXCL for SQ_CIPUT/SQ_CISVC queues. We leave SQ_EXCL on
7373 * until putnext, leavesq, or drain_syncq drops it.
7374 * That way we handle nested qwriter(INNER) without dropping
7375 * SQ_EXCL until the outermost qwriter callback routine is
7376 * done.
7377 */
7378 return;
7379 }
7380 SQ_PUTLOCKS_EXIT(sq);
7381 sqfill_events(sq, q, mp, func);
7382 }
7383
7384 /*
7385 * Synchronous callback support functions
7386 */
7387
7388 /*
7389 * Allocate a callback parameter structure.
7390 * Assumes that caller initializes the flags and the id.
7391 * Acquires SQLOCK(sq) if non-NULL is returned.
7392 */
7393 callbparams_t *
7394 callbparams_alloc(syncq_t *sq, void (*func)(void *), void *arg, int kmflags)
7395 {
7396 callbparams_t *cbp;
7397 size_t size = sizeof (callbparams_t);
7398
7399 cbp = kmem_alloc(size, kmflags & ~KM_PANIC);
7400
7401 /*
7402 * Only try tryhard allocation if the caller is ready to panic.
7403 * Otherwise just fail.
7404 */
7405 if (cbp == NULL) {
7406 if (kmflags & KM_PANIC)
7407 cbp = kmem_alloc_tryhard(sizeof (callbparams_t),
7408 &size, kmflags);
7409 else
7410 return (NULL);
7411 }
7412
7413 ASSERT(size >= sizeof (callbparams_t));
7414 cbp->cbp_size = size;
7415 cbp->cbp_sq = sq;
7416 cbp->cbp_func = func;
7417 cbp->cbp_arg = arg;
7418 mutex_enter(SQLOCK(sq));
7419 cbp->cbp_next = sq->sq_callbpend;
7420 sq->sq_callbpend = cbp;
7421 return (cbp);
7422 }
7423
7424 void
7425 callbparams_free(syncq_t *sq, callbparams_t *cbp)
7426 {
7427 callbparams_t **pp, *p;
7428
7429 ASSERT(MUTEX_HELD(SQLOCK(sq)));
7430
7431 for (pp = &sq->sq_callbpend; (p = *pp) != NULL; pp = &p->cbp_next) {
7432 if (p == cbp) {
7433 *pp = p->cbp_next;
7434 kmem_free(p, p->cbp_size);
7435 return;
7436 }
7437 }
7438 (void) (STRLOG(0, 0, 0, SL_CONSOLE,
7439 "callbparams_free: not found\n"));
7440 }
7441
7442 void
7443 callbparams_free_id(syncq_t *sq, callbparams_id_t id, int32_t flag)
7444 {
7445 callbparams_t **pp, *p;
7446
7447 ASSERT(MUTEX_HELD(SQLOCK(sq)));
7448
7449 for (pp = &sq->sq_callbpend; (p = *pp) != NULL; pp = &p->cbp_next) {
7450 if (p->cbp_id == id && p->cbp_flags == flag) {
7451 *pp = p->cbp_next;
7452 kmem_free(p, p->cbp_size);
7453 return;
7454 }
7455 }
7456 (void) (STRLOG(0, 0, 0, SL_CONSOLE,
7457 "callbparams_free_id: not found\n"));
7458 }
7459
7460 /*
7461 * Callback wrapper function used by once-only callbacks that can be
7462 * cancelled (qtimeout and qbufcall)
7463 * Contains inline version of entersq(sq, SQ_CALLBACK) that can be
7464 * cancelled by the qun* functions.
7465 */
7466 void
7467 qcallbwrapper(void *arg)
7468 {
7469 callbparams_t *cbp = arg;
7470 syncq_t *sq;
7471 uint16_t count = 0;
7472 uint16_t waitflags = SQ_STAYAWAY | SQ_EVENTS | SQ_EXCL;
7473 uint16_t type;
7474
7475 sq = cbp->cbp_sq;
7476 mutex_enter(SQLOCK(sq));
7477 type = sq->sq_type;
7478 if (!(type & SQ_CICB)) {
7479 count = sq->sq_count;
7480 SQ_PUTLOCKS_ENTER(sq);
7481 SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
7482 SUM_SQ_PUTCOUNTS(sq, count);
7483 sq->sq_needexcl++;
7484 ASSERT(sq->sq_needexcl != 0); /* wraparound */
7485 waitflags |= SQ_MESSAGES;
7486 }
7487 /* Can not handle exclusive entry at outer perimeter */
7488 ASSERT(type & SQ_COCB);
7489
7490 while ((sq->sq_flags & waitflags) || (!(type & SQ_CICB) &&count != 0)) {
7491 if ((sq->sq_callbflags & cbp->cbp_flags) &&
7492 (sq->sq_cancelid == cbp->cbp_id)) {
7493 /* timeout has been cancelled */
7494 sq->sq_callbflags |= SQ_CALLB_BYPASSED;
7495 callbparams_free(sq, cbp);
7496 if (!(type & SQ_CICB)) {
7497 ASSERT(sq->sq_needexcl > 0);
7498 sq->sq_needexcl--;
7499 if (sq->sq_needexcl == 0) {
7500 SQ_PUTCOUNT_SETFAST_LOCKED(sq);
7501 }
7502 SQ_PUTLOCKS_EXIT(sq);
7503 }
7504 mutex_exit(SQLOCK(sq));
7505 return;
7506 }
7507 sq->sq_flags |= SQ_WANTWAKEUP;
7508 if (!(type & SQ_CICB)) {
7509 SQ_PUTLOCKS_EXIT(sq);
7510 }
7511 cv_wait(&sq->sq_wait, SQLOCK(sq));
7512 if (!(type & SQ_CICB)) {
7513 count = sq->sq_count;
7514 SQ_PUTLOCKS_ENTER(sq);
7515 SUM_SQ_PUTCOUNTS(sq, count);
7516 }
7517 }
7518
7519 sq->sq_count++;
7520 ASSERT(sq->sq_count != 0); /* Wraparound */
7521 if (!(type & SQ_CICB)) {
7522 ASSERT(count == 0);
7523 sq->sq_flags |= SQ_EXCL;
7524 ASSERT(sq->sq_needexcl > 0);
7525 sq->sq_needexcl--;
7526 if (sq->sq_needexcl == 0) {
7527 SQ_PUTCOUNT_SETFAST_LOCKED(sq);
7528 }
7529 SQ_PUTLOCKS_EXIT(sq);
7530 }
7531
7532 mutex_exit(SQLOCK(sq));
7533
7534 cbp->cbp_func(cbp->cbp_arg);
7535
7536 /*
7537 * We drop the lock only for leavesq to re-acquire it.
7538 * Possible optimization is inline of leavesq.
7539 */
7540 mutex_enter(SQLOCK(sq));
7541 callbparams_free(sq, cbp);
7542 mutex_exit(SQLOCK(sq));
7543 leavesq(sq, SQ_CALLBACK);
7544 }
7545
7546 /*
7547 * No need to grab sq_putlocks here. See comment in strsubr.h that
7548 * explains when sq_putlocks are used.
7549 *
7550 * sq_count (or one of the sq_putcounts) has already been
7551 * decremented by the caller, and if SQ_QUEUED, we need to call
7552 * drain_syncq (the global syncq drain).
7553 * If putnext_tail is called with the SQ_EXCL bit set, we are in
7554 * one of two states, non-CIPUT perimeter, and we need to clear
7555 * it, or we went exclusive in the put procedure. In any case,
7556 * we want to clear the bit now, and it is probably easier to do
7557 * this at the beginning of this function (remember, we hold
7558 * the SQLOCK). Lastly, if there are other messages queued
7559 * on the syncq (and not for our destination), enable the syncq
7560 * for background work.
7561 */
7562
7563 /* ARGSUSED */
7564 void
7565 putnext_tail(syncq_t *sq, queue_t *qp, uint32_t passflags)
7566 {
7567 uint16_t flags = sq->sq_flags;
7568
7569 ASSERT(MUTEX_HELD(SQLOCK(sq)));
7570 ASSERT(MUTEX_NOT_HELD(QLOCK(qp)));
7571
7572 /* Clear SQ_EXCL if set in passflags */
7573 if (passflags & SQ_EXCL) {
7574 flags &= ~SQ_EXCL;
7575 }
7576 if (flags & SQ_WANTWAKEUP) {
7577 flags &= ~SQ_WANTWAKEUP;
7578 cv_broadcast(&sq->sq_wait);
7579 }
7580 if (flags & SQ_WANTEXWAKEUP) {
7581 flags &= ~SQ_WANTEXWAKEUP;
7582 cv_broadcast(&sq->sq_exitwait);
7583 }
7584 sq->sq_flags = flags;
7585
7586 /*
7587 * We have cleared SQ_EXCL if we were asked to, and started
7588 * the wakeup process for waiters. If there are no writers
7589 * then we need to drain the syncq if we were told to, or
7590 * enable the background thread to do it.
7591 */
7592 if (!(flags & (SQ_STAYAWAY|SQ_EXCL))) {
7593 if ((passflags & SQ_QUEUED) ||
7594 (sq->sq_svcflags & SQ_DISABLED)) {
7595 /* drain_syncq will take care of events in the list */
7596 drain_syncq(sq);
7597 return;
7598 } else if (flags & SQ_QUEUED) {
7599 sqenable(sq);
7600 }
7601 }
7602 /* Drop the SQLOCK on exit */
7603 mutex_exit(SQLOCK(sq));
7604 TRACE_3(TR_FAC_STREAMS_FR, TR_PUTNEXT_END,
7605 "putnext_end:(%p, %p, %p) done", NULL, qp, sq);
7606 }
7607
7608 void
7609 set_qend(queue_t *q)
7610 {
7611 mutex_enter(QLOCK(q));
7612 if (!O_SAMESTR(q))
7613 q->q_flag |= QEND;
7614 else
7615 q->q_flag &= ~QEND;
7616 mutex_exit(QLOCK(q));
7617 q = _OTHERQ(q);
7618 mutex_enter(QLOCK(q));
7619 if (!O_SAMESTR(q))
7620 q->q_flag |= QEND;
7621 else
7622 q->q_flag &= ~QEND;
7623 mutex_exit(QLOCK(q));
7624 }
7625
7626 /*
7627 * Set QFULL in next service procedure queue (that cares) if not already
7628 * set and if there are already more messages on the syncq than
7629 * sq_max_size. If sq_max_size is 0, no flow control will be asserted on
7630 * any syncq.
7631 *
7632 * The fq here is the next queue with a service procedure. This is where
7633 * we would fail canputnext, so this is where we need to set QFULL.
7634 * In the case when fq != q we need to take QLOCK(fq) to set QFULL flag.
7635 *
7636 * We already have QLOCK at this point. To avoid cross-locks with
7637 * freezestr() which grabs all QLOCKs and with strlock() which grabs both
7638 * SQLOCK and sd_reflock, we need to drop respective locks first.
7639 */
7640 void
7641 set_qfull(queue_t *q)
7642 {
7643 queue_t *fq = NULL;
7644
7645 ASSERT(MUTEX_HELD(QLOCK(q)));
7646 if ((sq_max_size != 0) && (!(q->q_nfsrv->q_flag & QFULL)) &&
7647 (q->q_syncqmsgs > sq_max_size)) {
7648 if ((fq = q->q_nfsrv) == q) {
7649 fq->q_flag |= QFULL;
7650 } else {
7651 mutex_exit(QLOCK(q));
7652 mutex_enter(QLOCK(fq));
7653 fq->q_flag |= QFULL;
7654 mutex_exit(QLOCK(fq));
7655 mutex_enter(QLOCK(q));
7656 }
7657 }
7658 }
7659
7660 void
7661 clr_qfull(queue_t *q)
7662 {
7663 queue_t *oq = q;
7664
7665 q = q->q_nfsrv;
7666 /* Fast check if there is any work to do before getting the lock. */
7667 if ((q->q_flag & (QFULL|QWANTW)) == 0) {
7668 return;
7669 }
7670
7671 /*
7672 * Do not reset QFULL (and backenable) if the q_count is the reason
7673 * for QFULL being set.
7674 */
7675 mutex_enter(QLOCK(q));
7676 /*
7677 * If queue is empty i.e q_mblkcnt is zero, queue can not be full.
7678 * Hence clear the QFULL.
7679 * If both q_count and q_mblkcnt are less than the hiwat mark,
7680 * clear the QFULL.
7681 */
7682 if (q->q_mblkcnt == 0 || ((q->q_count < q->q_hiwat) &&
7683 (q->q_mblkcnt < q->q_hiwat))) {
7684 q->q_flag &= ~QFULL;
7685 /*
7686 * A little more confusing, how about this way:
7687 * if someone wants to write,
7688 * AND
7689 * both counts are less than the lowat mark
7690 * OR
7691 * the lowat mark is zero
7692 * THEN
7693 * backenable
7694 */
7695 if ((q->q_flag & QWANTW) &&
7696 (((q->q_count < q->q_lowat) &&
7697 (q->q_mblkcnt < q->q_lowat)) || q->q_lowat == 0)) {
7698 q->q_flag &= ~QWANTW;
7699 mutex_exit(QLOCK(q));
7700 backenable(oq, 0);
7701 } else
7702 mutex_exit(QLOCK(q));
7703 } else
7704 mutex_exit(QLOCK(q));
7705 }
7706
7707 /*
7708 * Set the forward service procedure pointer.
7709 *
7710 * Called at insert-time to cache a queue's next forward service procedure in
7711 * q_nfsrv; used by canput() and canputnext(). If the queue to be inserted
7712 * has a service procedure then q_nfsrv points to itself. If the queue to be
7713 * inserted does not have a service procedure, then q_nfsrv points to the next
7714 * queue forward that has a service procedure. If the queue is at the logical
7715 * end of the stream (driver for write side, stream head for the read side)
7716 * and does not have a service procedure, then q_nfsrv also points to itself.
7717 */
7718 void
7719 set_nfsrv_ptr(
7720 queue_t *rnew, /* read queue pointer to new module */
7721 queue_t *wnew, /* write queue pointer to new module */
7722 queue_t *prev_rq, /* read queue pointer to the module above */
7723 queue_t *prev_wq) /* write queue pointer to the module above */
7724 {
7725 queue_t *qp;
7726
7727 if (prev_wq->q_next == NULL) {
7728 /*
7729 * Insert the driver, initialize the driver and stream head.
7730 * In this case, prev_rq/prev_wq should be the stream head.
7731 * _I_INSERT does not allow inserting a driver. Make sure
7732 * that it is not an insertion.
7733 */
7734 ASSERT(!(rnew->q_flag & _QINSERTING));
7735 wnew->q_nfsrv = wnew;
7736 if (rnew->q_qinfo->qi_srvp)
7737 rnew->q_nfsrv = rnew;
7738 else
7739 rnew->q_nfsrv = prev_rq;
7740 prev_rq->q_nfsrv = prev_rq;
7741 prev_wq->q_nfsrv = prev_wq;
7742 } else {
7743 /*
7744 * set up read side q_nfsrv pointer. This MUST be done
7745 * before setting the write side, because the setting of
7746 * the write side for a fifo may depend on it.
7747 *
7748 * Suppose we have a fifo that only has pipemod pushed.
7749 * pipemod has no read or write service procedures, so
7750 * nfsrv for both pipemod queues points to prev_rq (the
7751 * stream read head). Now push bufmod (which has only a
7752 * read service procedure). Doing the write side first,
7753 * wnew->q_nfsrv is set to pipemod's writeq nfsrv, which
7754 * is WRONG; the next queue forward from wnew with a
7755 * service procedure will be rnew, not the stream read head.
7756 * Since the downstream queue (which in the case of a fifo
7757 * is the read queue rnew) can affect upstream queues, it
7758 * needs to be done first. Setting up the read side first
7759 * sets nfsrv for both pipemod queues to rnew and then
7760 * when the write side is set up, wnew-q_nfsrv will also
7761 * point to rnew.
7762 */
7763 if (rnew->q_qinfo->qi_srvp) {
7764 /*
7765 * use _OTHERQ() because, if this is a pipe, next
7766 * module may have been pushed from other end and
7767 * q_next could be a read queue.
7768 */
7769 qp = _OTHERQ(prev_wq->q_next);
7770 while (qp && qp->q_nfsrv != qp) {
7771 qp->q_nfsrv = rnew;
7772 qp = backq(qp);
7773 }
7774 rnew->q_nfsrv = rnew;
7775 } else
7776 rnew->q_nfsrv = prev_rq->q_nfsrv;
7777
7778 /* set up write side q_nfsrv pointer */
7779 if (wnew->q_qinfo->qi_srvp) {
7780 wnew->q_nfsrv = wnew;
7781
7782 /*
7783 * For insertion, need to update nfsrv of the modules
7784 * above which do not have a service routine.
7785 */
7786 if (rnew->q_flag & _QINSERTING) {
7787 for (qp = prev_wq;
7788 qp != NULL && qp->q_nfsrv != qp;
7789 qp = backq(qp)) {
7790 qp->q_nfsrv = wnew->q_nfsrv;
7791 }
7792 }
7793 } else {
7794 if (prev_wq->q_next == prev_rq)
7795 /*
7796 * Since prev_wq/prev_rq are the middle of a
7797 * fifo, wnew/rnew will also be the middle of
7798 * a fifo and wnew's nfsrv is same as rnew's.
7799 */
7800 wnew->q_nfsrv = rnew->q_nfsrv;
7801 else
7802 wnew->q_nfsrv = prev_wq->q_next->q_nfsrv;
7803 }
7804 }
7805 }
7806
7807 /*
7808 * Reset the forward service procedure pointer; called at remove-time.
7809 */
7810 void
7811 reset_nfsrv_ptr(queue_t *rqp, queue_t *wqp)
7812 {
7813 queue_t *tmp_qp;
7814
7815 /* Reset the write side q_nfsrv pointer for _I_REMOVE */
7816 if ((rqp->q_flag & _QREMOVING) && (wqp->q_qinfo->qi_srvp != NULL)) {
7817 for (tmp_qp = backq(wqp);
7818 tmp_qp != NULL && tmp_qp->q_nfsrv == wqp;
7819 tmp_qp = backq(tmp_qp)) {
7820 tmp_qp->q_nfsrv = wqp->q_nfsrv;
7821 }
7822 }
7823
7824 /* reset the read side q_nfsrv pointer */
7825 if (rqp->q_qinfo->qi_srvp) {
7826 if (wqp->q_next) { /* non-driver case */
7827 tmp_qp = _OTHERQ(wqp->q_next);
7828 while (tmp_qp && tmp_qp->q_nfsrv == rqp) {
7829 /* Note that rqp->q_next cannot be NULL */
7830 ASSERT(rqp->q_next != NULL);
7831 tmp_qp->q_nfsrv = rqp->q_next->q_nfsrv;
7832 tmp_qp = backq(tmp_qp);
7833 }
7834 }
7835 }
7836 }
7837
7838 /*
7839 * This routine should be called after all stream geometry changes to update
7840 * the stream head cached struio() rd/wr queue pointers. Note must be called
7841 * with the streamlock()ed.
7842 *
7843 * Note: only enables Synchronous STREAMS for a side of a Stream which has
7844 * an explicit synchronous barrier module queue. That is, a queue that
7845 * has specified a struio() type.
7846 */
7847 static void
7848 strsetuio(stdata_t *stp)
7849 {
7850 queue_t *wrq;
7851
7852 if (stp->sd_flag & STPLEX) {
7853 /*
7854 * Not streamhead, but a mux, so no Synchronous STREAMS.
7855 */
7856 stp->sd_struiowrq = NULL;
7857 stp->sd_struiordq = NULL;
7858 return;
7859 }
7860 /*
7861 * Scan the write queue(s) while synchronous
7862 * until we find a qinfo uio type specified.
7863 */
7864 wrq = stp->sd_wrq->q_next;
7865 while (wrq) {
7866 if (wrq->q_struiot == STRUIOT_NONE) {
7867 wrq = 0;
7868 break;
7869 }
7870 if (wrq->q_struiot != STRUIOT_DONTCARE)
7871 break;
7872 if (! _SAMESTR(wrq)) {
7873 wrq = 0;
7874 break;
7875 }
7876 wrq = wrq->q_next;
7877 }
7878 stp->sd_struiowrq = wrq;
7879 /*
7880 * Scan the read queue(s) while synchronous
7881 * until we find a qinfo uio type specified.
7882 */
7883 wrq = stp->sd_wrq->q_next;
7884 while (wrq) {
7885 if (_RD(wrq)->q_struiot == STRUIOT_NONE) {
7886 wrq = 0;
7887 break;
7888 }
7889 if (_RD(wrq)->q_struiot != STRUIOT_DONTCARE)
7890 break;
7891 if (! _SAMESTR(wrq)) {
7892 wrq = 0;
7893 break;
7894 }
7895 wrq = wrq->q_next;
7896 }
7897 stp->sd_struiordq = wrq ? _RD(wrq) : 0;
7898 }
7899
7900 /*
7901 * pass_wput, unblocks the passthru queues, so that
7902 * messages can arrive at muxs lower read queue, before
7903 * I_LINK/I_UNLINK is acked/nacked.
7904 */
7905 static void
7906 pass_wput(queue_t *q, mblk_t *mp)
7907 {
7908 syncq_t *sq;
7909
7910 sq = _RD(q)->q_syncq;
7911 if (sq->sq_flags & SQ_BLOCKED)
7912 unblocksq(sq, SQ_BLOCKED, 0);
7913 putnext(q, mp);
7914 }
7915
7916 /*
7917 * Set up queues for the link/unlink.
7918 * Create a new queue and block it and then insert it
7919 * below the stream head on the lower stream.
7920 * This prevents any messages from arriving during the setq
7921 * as well as while the mux is processing the LINK/I_UNLINK.
7922 * The blocked passq is unblocked once the LINK/I_UNLINK has
7923 * been acked or nacked or if a message is generated and sent
7924 * down muxs write put procedure.
7925 * See pass_wput().
7926 *
7927 * After the new queue is inserted, all messages coming from below are
7928 * blocked. The call to strlock will ensure that all activity in the stream head
7929 * read queue syncq is stopped (sq_count drops to zero).
7930 */
7931 static queue_t *
7932 link_addpassthru(stdata_t *stpdown)
7933 {
7934 queue_t *passq;
7935 sqlist_t sqlist;
7936
7937 passq = allocq();
7938 STREAM(passq) = STREAM(_WR(passq)) = stpdown;
7939 /* setq might sleep in allocator - avoid holding locks. */
7940 setq(passq, &passthru_rinit, &passthru_winit, NULL, QPERQ,
7941 SQ_CI|SQ_CO, B_FALSE);
7942 claimq(passq);
7943 blocksq(passq->q_syncq, SQ_BLOCKED, 1);
7944 insertq(STREAM(passq), passq);
7945
7946 /*
7947 * Use strlock() to wait for the stream head sq_count to drop to zero
7948 * since we are going to change q_ptr in the stream head. Note that
7949 * insertq() doesn't wait for any syncq counts to drop to zero.
7950 */
7951 sqlist.sqlist_head = NULL;
7952 sqlist.sqlist_index = 0;
7953 sqlist.sqlist_size = sizeof (sqlist_t);
7954 sqlist_insert(&sqlist, _RD(stpdown->sd_wrq)->q_syncq);
7955 strlock(stpdown, &sqlist);
7956 strunlock(stpdown, &sqlist);
7957
7958 releaseq(passq);
7959 return (passq);
7960 }
7961
7962 /*
7963 * Let messages flow up into the mux by removing
7964 * the passq.
7965 */
7966 static void
7967 link_rempassthru(queue_t *passq)
7968 {
7969 claimq(passq);
7970 removeq(passq);
7971 releaseq(passq);
7972 freeq(passq);
7973 }
7974
7975 /*
7976 * Wait for the condition variable pointed to by `cvp' to be signaled,
7977 * or for `tim' milliseconds to elapse, whichever comes first. If `tim'
7978 * is negative, then there is no time limit. If `nosigs' is non-zero,
7979 * then the wait will be non-interruptible.
7980 *
7981 * Returns >0 if signaled, 0 if interrupted, or -1 upon timeout.
7982 */
7983 clock_t
7984 str_cv_wait(kcondvar_t *cvp, kmutex_t *mp, clock_t tim, int nosigs)
7985 {
7986 clock_t ret;
7987
7988 if (tim < 0) {
7989 if (nosigs) {
7990 cv_wait(cvp, mp);
7991 ret = 1;
7992 } else {
7993 ret = cv_wait_sig(cvp, mp);
7994 }
7995 } else if (tim > 0) {
7996 /*
7997 * convert milliseconds to clock ticks
7998 */
7999 if (nosigs) {
8000 ret = cv_reltimedwait(cvp, mp,
8001 MSEC_TO_TICK_ROUNDUP(tim), TR_CLOCK_TICK);
8002 } else {
8003 ret = cv_reltimedwait_sig(cvp, mp,
8004 MSEC_TO_TICK_ROUNDUP(tim), TR_CLOCK_TICK);
8005 }
8006 } else {
8007 ret = -1;
8008 }
8009 return (ret);
8010 }
8011
8012 /*
8013 * Wait until the stream head can determine if it is at the mark but
8014 * don't wait forever to prevent a race condition between the "mark" state
8015 * in the stream head and any mark state in the caller/user of this routine.
8016 *
8017 * This is used by sockets and for a socket it would be incorrect
8018 * to return a failure for SIOCATMARK when there is no data in the receive
8019 * queue and the marked urgent data is traveling up the stream.
8020 *
8021 * This routine waits until the mark is known by waiting for one of these
8022 * three events:
8023 * The stream head read queue becoming non-empty (including an EOF).
8024 * The STRATMARK flag being set (due to a MSGMARKNEXT message).
8025 * The STRNOTATMARK flag being set (which indicates that the transport
8026 * has sent a MSGNOTMARKNEXT message to indicate that it is not at
8027 * the mark).
8028 *
8029 * The routine returns 1 if the stream is at the mark; 0 if it can
8030 * be determined that the stream is not at the mark.
8031 * If the wait times out and it can't determine
8032 * whether or not the stream might be at the mark the routine will return -1.
8033 *
8034 * Note: This routine should only be used when a mark is pending i.e.,
8035 * in the socket case the SIGURG has been posted.
8036 * Note2: This can not wakeup just because synchronous streams indicate
8037 * that data is available since it is not possible to use the synchronous
8038 * streams interfaces to determine the b_flag value for the data queued below
8039 * the stream head.
8040 */
8041 int
8042 strwaitmark(vnode_t *vp)
8043 {
8044 struct stdata *stp = vp->v_stream;
8045 queue_t *rq = _RD(stp->sd_wrq);
8046 int mark;
8047
8048 mutex_enter(&stp->sd_lock);
8049 while (rq->q_first == NULL &&
8050 !(stp->sd_flag & (STRATMARK|STRNOTATMARK|STREOF))) {
8051 stp->sd_flag |= RSLEEP;
8052
8053 /* Wait for 100 milliseconds for any state change. */
8054 if (str_cv_wait(&rq->q_wait, &stp->sd_lock, 100, 1) == -1) {
8055 mutex_exit(&stp->sd_lock);
8056 return (-1);
8057 }
8058 }
8059 if (stp->sd_flag & STRATMARK)
8060 mark = 1;
8061 else if (rq->q_first != NULL && (rq->q_first->b_flag & MSGMARK))
8062 mark = 1;
8063 else
8064 mark = 0;
8065
8066 mutex_exit(&stp->sd_lock);
8067 return (mark);
8068 }
8069
8070 /*
8071 * Set a read side error. If persist is set change the socket error
8072 * to persistent. If errfunc is set install the function as the exported
8073 * error handler.
8074 */
8075 void
8076 strsetrerror(vnode_t *vp, int error, int persist, errfunc_t errfunc)
8077 {
8078 struct stdata *stp = vp->v_stream;
8079
8080 mutex_enter(&stp->sd_lock);
8081 stp->sd_rerror = error;
8082 if (error == 0 && errfunc == NULL)
8083 stp->sd_flag &= ~STRDERR;
8084 else
8085 stp->sd_flag |= STRDERR;
8086 if (persist) {
8087 stp->sd_flag &= ~STRDERRNONPERSIST;
8088 } else {
8089 stp->sd_flag |= STRDERRNONPERSIST;
8090 }
8091 stp->sd_rderrfunc = errfunc;
8092 if (error != 0 || errfunc != NULL) {
8093 cv_broadcast(&_RD(stp->sd_wrq)->q_wait); /* readers */
8094 cv_broadcast(&stp->sd_wrq->q_wait); /* writers */
8095 cv_broadcast(&stp->sd_monitor); /* ioctllers */
8096
8097 mutex_exit(&stp->sd_lock);
8098 pollwakeup(&stp->sd_pollist, POLLERR);
8099 mutex_enter(&stp->sd_lock);
8100
8101 if (stp->sd_sigflags & S_ERROR)
8102 strsendsig(stp->sd_siglist, S_ERROR, 0, error);
8103 }
8104 mutex_exit(&stp->sd_lock);
8105 }
8106
8107 /*
8108 * Set a write side error. If persist is set change the socket error
8109 * to persistent.
8110 */
8111 void
8112 strsetwerror(vnode_t *vp, int error, int persist, errfunc_t errfunc)
8113 {
8114 struct stdata *stp = vp->v_stream;
8115
8116 mutex_enter(&stp->sd_lock);
8117 stp->sd_werror = error;
8118 if (error == 0 && errfunc == NULL)
8119 stp->sd_flag &= ~STWRERR;
8120 else
8121 stp->sd_flag |= STWRERR;
8122 if (persist) {
8123 stp->sd_flag &= ~STWRERRNONPERSIST;
8124 } else {
8125 stp->sd_flag |= STWRERRNONPERSIST;
8126 }
8127 stp->sd_wrerrfunc = errfunc;
8128 if (error != 0 || errfunc != NULL) {
8129 cv_broadcast(&_RD(stp->sd_wrq)->q_wait); /* readers */
8130 cv_broadcast(&stp->sd_wrq->q_wait); /* writers */
8131 cv_broadcast(&stp->sd_monitor); /* ioctllers */
8132
8133 mutex_exit(&stp->sd_lock);
8134 pollwakeup(&stp->sd_pollist, POLLERR);
8135 mutex_enter(&stp->sd_lock);
8136
8137 if (stp->sd_sigflags & S_ERROR)
8138 strsendsig(stp->sd_siglist, S_ERROR, 0, error);
8139 }
8140 mutex_exit(&stp->sd_lock);
8141 }
8142
8143 /*
8144 * Make the stream return 0 (EOF) when all data has been read.
8145 * No effect on write side.
8146 */
8147 void
8148 strseteof(vnode_t *vp, int eof)
8149 {
8150 struct stdata *stp = vp->v_stream;
8151
8152 mutex_enter(&stp->sd_lock);
8153 if (!eof) {
8154 stp->sd_flag &= ~STREOF;
8155 mutex_exit(&stp->sd_lock);
8156 return;
8157 }
8158 stp->sd_flag |= STREOF;
8159 if (stp->sd_flag & RSLEEP) {
8160 stp->sd_flag &= ~RSLEEP;
8161 cv_broadcast(&_RD(stp->sd_wrq)->q_wait);
8162 }
8163
8164 mutex_exit(&stp->sd_lock);
8165 pollwakeup(&stp->sd_pollist, POLLIN|POLLRDNORM);
8166 mutex_enter(&stp->sd_lock);
8167
8168 if (stp->sd_sigflags & (S_INPUT|S_RDNORM))
8169 strsendsig(stp->sd_siglist, S_INPUT|S_RDNORM, 0, 0);
8170 mutex_exit(&stp->sd_lock);
8171 }
8172
8173 void
8174 strflushrq(vnode_t *vp, int flag)
8175 {
8176 struct stdata *stp = vp->v_stream;
8177
8178 mutex_enter(&stp->sd_lock);
8179 flushq(_RD(stp->sd_wrq), flag);
8180 mutex_exit(&stp->sd_lock);
8181 }
8182
8183 void
8184 strsetrputhooks(vnode_t *vp, uint_t flags,
8185 msgfunc_t protofunc, msgfunc_t miscfunc)
8186 {
8187 struct stdata *stp = vp->v_stream;
8188
8189 mutex_enter(&stp->sd_lock);
8190
8191 if (protofunc == NULL)
8192 stp->sd_rprotofunc = strrput_proto;
8193 else
8194 stp->sd_rprotofunc = protofunc;
8195
8196 if (miscfunc == NULL)
8197 stp->sd_rmiscfunc = strrput_misc;
8198 else
8199 stp->sd_rmiscfunc = miscfunc;
8200
8201 if (flags & SH_CONSOL_DATA)
8202 stp->sd_rput_opt |= SR_CONSOL_DATA;
8203 else
8204 stp->sd_rput_opt &= ~SR_CONSOL_DATA;
8205
8206 if (flags & SH_SIGALLDATA)
8207 stp->sd_rput_opt |= SR_SIGALLDATA;
8208 else
8209 stp->sd_rput_opt &= ~SR_SIGALLDATA;
8210
8211 if (flags & SH_IGN_ZEROLEN)
8212 stp->sd_rput_opt |= SR_IGN_ZEROLEN;
8213 else
8214 stp->sd_rput_opt &= ~SR_IGN_ZEROLEN;
8215
8216 mutex_exit(&stp->sd_lock);
8217 }
8218
8219 void
8220 strsetwputhooks(vnode_t *vp, uint_t flags, clock_t closetime)
8221 {
8222 struct stdata *stp = vp->v_stream;
8223
8224 mutex_enter(&stp->sd_lock);
8225 stp->sd_closetime = closetime;
8226
8227 if (flags & SH_SIGPIPE)
8228 stp->sd_wput_opt |= SW_SIGPIPE;
8229 else
8230 stp->sd_wput_opt &= ~SW_SIGPIPE;
8231 if (flags & SH_RECHECK_ERR)
8232 stp->sd_wput_opt |= SW_RECHECK_ERR;
8233 else
8234 stp->sd_wput_opt &= ~SW_RECHECK_ERR;
8235
8236 mutex_exit(&stp->sd_lock);
8237 }
8238
8239 void
8240 strsetrwputdatahooks(vnode_t *vp, msgfunc_t rdatafunc, msgfunc_t wdatafunc)
8241 {
8242 struct stdata *stp = vp->v_stream;
8243
8244 mutex_enter(&stp->sd_lock);
8245
8246 stp->sd_rputdatafunc = rdatafunc;
8247 stp->sd_wputdatafunc = wdatafunc;
8248
8249 mutex_exit(&stp->sd_lock);
8250 }
8251
8252 /* Used within framework when the queue is already locked */
8253 void
8254 qenable_locked(queue_t *q)
8255 {
8256 stdata_t *stp = STREAM(q);
8257
8258 ASSERT(MUTEX_HELD(QLOCK(q)));
8259
8260 if (!q->q_qinfo->qi_srvp)
8261 return;
8262
8263 /*
8264 * Do not place on run queue if already enabled or closing.
8265 */
8266 if (q->q_flag & (QWCLOSE|QENAB))
8267 return;
8268
8269 /*
8270 * mark queue enabled and place on run list if it is not already being
8271 * serviced. If it is serviced, the runservice() function will detect
8272 * that QENAB is set and call service procedure before clearing
8273 * QINSERVICE flag.
8274 */
8275 q->q_flag |= QENAB;
8276 if (q->q_flag & QINSERVICE)
8277 return;
8278
8279 /* Record the time of qenable */
8280 q->q_qtstamp = ddi_get_lbolt();
8281
8282 /*
8283 * Put the queue in the stp list and schedule it for background
8284 * processing if it is not already scheduled or if stream head does not
8285 * intent to process it in the foreground later by setting
8286 * STRS_WILLSERVICE flag.
8287 */
8288 mutex_enter(&stp->sd_qlock);
8289 /*
8290 * If there are already something on the list, stp flags should show
8291 * intention to drain it.
8292 */
8293 IMPLY(STREAM_NEEDSERVICE(stp),
8294 (stp->sd_svcflags & (STRS_WILLSERVICE | STRS_SCHEDULED)));
8295
8296 ENQUEUE(q, stp->sd_qhead, stp->sd_qtail, q_link);
8297 stp->sd_nqueues++;
8298
8299 /*
8300 * If no one will drain this stream we are the first producer and
8301 * need to schedule it for background thread.
8302 */
8303 if (!(stp->sd_svcflags & (STRS_WILLSERVICE | STRS_SCHEDULED))) {
8304 /*
8305 * No one will service this stream later, so we have to
8306 * schedule it now.
8307 */
8308 STRSTAT(stenables);
8309 stp->sd_svcflags |= STRS_SCHEDULED;
8310 stp->sd_servid = (void *)taskq_dispatch(streams_taskq,
8311 (task_func_t *)stream_service, stp, TQ_NOSLEEP|TQ_NOQUEUE);
8312
8313 if (stp->sd_servid == NULL) {
8314 /*
8315 * Task queue failed so fail over to the backup
8316 * servicing thread.
8317 */
8318 STRSTAT(taskqfails);
8319 /*
8320 * It is safe to clear STRS_SCHEDULED flag because it
8321 * was set by this thread above.
8322 */
8323 stp->sd_svcflags &= ~STRS_SCHEDULED;
8324
8325 /*
8326 * Failover scheduling is protected by service_queue
8327 * lock.
8328 */
8329 mutex_enter(&service_queue);
8330 ASSERT((stp->sd_qhead == q) && (stp->sd_qtail == q));
8331 ASSERT(q->q_link == NULL);
8332 /*
8333 * Append the queue to qhead/qtail list.
8334 */
8335 if (qhead == NULL)
8336 qhead = q;
8337 else
8338 qtail->q_link = q;
8339 qtail = q;
8340 /*
8341 * Clear stp queue list.
8342 */
8343 stp->sd_qhead = stp->sd_qtail = NULL;
8344 stp->sd_nqueues = 0;
8345 /*
8346 * Wakeup background queue processing thread.
8347 */
8348 cv_signal(&services_to_run);
8349 mutex_exit(&service_queue);
8350 }
8351 }
8352 mutex_exit(&stp->sd_qlock);
8353 }
8354
8355 static void
8356 queue_service(queue_t *q)
8357 {
8358 /*
8359 * The queue in the list should have
8360 * QENAB flag set and should not have
8361 * QINSERVICE flag set. QINSERVICE is
8362 * set when the queue is dequeued and
8363 * qenable_locked doesn't enqueue a
8364 * queue with QINSERVICE set.
8365 */
8366
8367 ASSERT(!(q->q_flag & QINSERVICE));
8368 ASSERT((q->q_flag & QENAB));
8369 mutex_enter(QLOCK(q));
8370 q->q_flag &= ~QENAB;
8371 q->q_flag |= QINSERVICE;
8372 mutex_exit(QLOCK(q));
8373 runservice(q);
8374 }
8375
8376 static void
8377 syncq_service(syncq_t *sq)
8378 {
8379 STRSTAT(syncqservice);
8380 mutex_enter(SQLOCK(sq));
8381 ASSERT(!(sq->sq_svcflags & SQ_SERVICE));
8382 ASSERT(sq->sq_servcount != 0);
8383 ASSERT(sq->sq_next == NULL);
8384
8385 /* if we came here from the background thread, clear the flag */
8386 if (sq->sq_svcflags & SQ_BGTHREAD)
8387 sq->sq_svcflags &= ~SQ_BGTHREAD;
8388
8389 /* let drain_syncq know that it's being called in the background */
8390 sq->sq_svcflags |= SQ_SERVICE;
8391 drain_syncq(sq);
8392 }
8393
8394 static void
8395 qwriter_outer_service(syncq_t *outer)
8396 {
8397 /*
8398 * Note that SQ_WRITER is used on the outer perimeter
8399 * to signal that a qwriter(OUTER) is either investigating
8400 * running or that it is actually running a function.
8401 */
8402 outer_enter(outer, SQ_BLOCKED|SQ_WRITER);
8403
8404 /*
8405 * All inner syncq are empty and have SQ_WRITER set
8406 * to block entering the outer perimeter.
8407 *
8408 * We do not need to explicitly call write_now since
8409 * outer_exit does it for us.
8410 */
8411 outer_exit(outer);
8412 }
8413
8414 static void
8415 mblk_free(mblk_t *mp)
8416 {
8417 dblk_t *dbp = mp->b_datap;
8418 frtn_t *frp = dbp->db_frtnp;
8419
8420 mp->b_next = NULL;
8421 if (dbp->db_fthdr != NULL)
8422 str_ftfree(dbp);
8423
8424 ASSERT(dbp->db_fthdr == NULL);
8425 frp->free_func(frp->free_arg);
8426 ASSERT(dbp->db_mblk == mp);
8427
8428 if (dbp->db_credp != NULL) {
8429 crfree(dbp->db_credp);
8430 dbp->db_credp = NULL;
8431 }
8432 dbp->db_cpid = -1;
8433 dbp->db_struioflag = 0;
8434 dbp->db_struioun.cksum.flags = 0;
8435
8436 kmem_cache_free(dbp->db_cache, dbp);
8437 }
8438
8439 /*
8440 * Background processing of the stream queue list.
8441 */
8442 static void
8443 stream_service(stdata_t *stp)
8444 {
8445 queue_t *q;
8446
8447 mutex_enter(&stp->sd_qlock);
8448
8449 STR_SERVICE(stp, q);
8450
8451 stp->sd_svcflags &= ~STRS_SCHEDULED;
8452 stp->sd_servid = NULL;
8453 cv_signal(&stp->sd_qcv);
8454 mutex_exit(&stp->sd_qlock);
8455 }
8456
8457 /*
8458 * Foreground processing of the stream queue list.
8459 */
8460 void
8461 stream_runservice(stdata_t *stp)
8462 {
8463 queue_t *q;
8464
8465 mutex_enter(&stp->sd_qlock);
8466 STRSTAT(rservice);
8467 /*
8468 * We are going to drain this stream queue list, so qenable_locked will
8469 * not schedule it until we finish.
8470 */
8471 stp->sd_svcflags |= STRS_WILLSERVICE;
8472
8473 STR_SERVICE(stp, q);
8474
8475 stp->sd_svcflags &= ~STRS_WILLSERVICE;
8476 mutex_exit(&stp->sd_qlock);
8477 /*
8478 * Help backup background thread to drain the qhead/qtail list.
8479 */
8480 while (qhead != NULL) {
8481 STRSTAT(qhelps);
8482 mutex_enter(&service_queue);
8483 DQ(q, qhead, qtail, q_link);
8484 mutex_exit(&service_queue);
8485 if (q != NULL)
8486 queue_service(q);
8487 }
8488 }
8489
8490 void
8491 stream_willservice(stdata_t *stp)
8492 {
8493 mutex_enter(&stp->sd_qlock);
8494 stp->sd_svcflags |= STRS_WILLSERVICE;
8495 mutex_exit(&stp->sd_qlock);
8496 }
8497
8498 /*
8499 * Replace the cred currently in the mblk with a different one.
8500 * Also update db_cpid.
8501 */
8502 void
8503 mblk_setcred(mblk_t *mp, cred_t *cr, pid_t cpid)
8504 {
8505 dblk_t *dbp = mp->b_datap;
8506 cred_t *ocr = dbp->db_credp;
8507
8508 ASSERT(cr != NULL);
8509
8510 if (cr != ocr) {
8511 crhold(dbp->db_credp = cr);
8512 if (ocr != NULL)
8513 crfree(ocr);
8514 }
8515 /* Don't overwrite with NOPID */
8516 if (cpid != NOPID)
8517 dbp->db_cpid = cpid;
8518 }
8519
8520 /*
8521 * If the src message has a cred, then replace the cred currently in the mblk
8522 * with it.
8523 * Also update db_cpid.
8524 */
8525 void
8526 mblk_copycred(mblk_t *mp, const mblk_t *src)
8527 {
8528 dblk_t *dbp = mp->b_datap;
8529 cred_t *cr, *ocr;
8530 pid_t cpid;
8531
8532 cr = msg_getcred(src, &cpid);
8533 if (cr == NULL)
8534 return;
8535
8536 ocr = dbp->db_credp;
8537 if (cr != ocr) {
8538 crhold(dbp->db_credp = cr);
8539 if (ocr != NULL)
8540 crfree(ocr);
8541 }
8542 /* Don't overwrite with NOPID */
8543 if (cpid != NOPID)
8544 dbp->db_cpid = cpid;
8545 }
8546
8547 int
8548 hcksum_assoc(mblk_t *mp, multidata_t *mmd, pdesc_t *pd,
8549 uint32_t start, uint32_t stuff, uint32_t end, uint32_t value,
8550 uint32_t flags, int km_flags)
8551 {
8552 int rc = 0;
8553
8554 ASSERT(DB_TYPE(mp) == M_DATA || DB_TYPE(mp) == M_MULTIDATA);
8555 if (mp->b_datap->db_type == M_DATA) {
8556 /* Associate values for M_DATA type */
8557 DB_CKSUMSTART(mp) = (intptr_t)start;
8558 DB_CKSUMSTUFF(mp) = (intptr_t)stuff;
8559 DB_CKSUMEND(mp) = (intptr_t)end;
8560 DB_CKSUMFLAGS(mp) = flags;
8561 DB_CKSUM16(mp) = (uint16_t)value;
8562
8563 } else {
8564 pattrinfo_t pa_info;
8565
8566 ASSERT(mmd != NULL);
8567
8568 pa_info.type = PATTR_HCKSUM;
8569 pa_info.len = sizeof (pattr_hcksum_t);
8570
8571 if (mmd_addpattr(mmd, pd, &pa_info, B_TRUE, km_flags) != NULL) {
8572 pattr_hcksum_t *hck = (pattr_hcksum_t *)pa_info.buf;
8573
8574 hck->hcksum_start_offset = start;
8575 hck->hcksum_stuff_offset = stuff;
8576 hck->hcksum_end_offset = end;
8577 hck->hcksum_cksum_val.inet_cksum = (uint16_t)value;
8578 hck->hcksum_flags = flags;
8579 } else {
8580 rc = -1;
8581 }
8582 }
8583 return (rc);
8584 }
8585
8586 void
8587 hcksum_retrieve(mblk_t *mp, multidata_t *mmd, pdesc_t *pd,
8588 uint32_t *start, uint32_t *stuff, uint32_t *end,
8589 uint32_t *value, uint32_t *flags)
8590 {
8591 ASSERT(DB_TYPE(mp) == M_DATA || DB_TYPE(mp) == M_MULTIDATA);
8592 if (mp->b_datap->db_type == M_DATA) {
8593 if (flags != NULL) {
8594 *flags = DB_CKSUMFLAGS(mp) & HCK_FLAGS;
8595 if ((*flags & (HCK_PARTIALCKSUM |
8596 HCK_FULLCKSUM)) != 0) {
8597 if (value != NULL)
8598 *value = (uint32_t)DB_CKSUM16(mp);
8599 if ((*flags & HCK_PARTIALCKSUM) != 0) {
8600 if (start != NULL)
8601 *start =
8602 (uint32_t)DB_CKSUMSTART(mp);
8603 if (stuff != NULL)
8604 *stuff =
8605 (uint32_t)DB_CKSUMSTUFF(mp);
8606 if (end != NULL)
8607 *end =
8608 (uint32_t)DB_CKSUMEND(mp);
8609 }
8610 }
8611 }
8612 } else {
8613 pattrinfo_t hck_attr = {PATTR_HCKSUM};
8614
8615 ASSERT(mmd != NULL);
8616
8617 /* get hardware checksum attribute */
8618 if (mmd_getpattr(mmd, pd, &hck_attr) != NULL) {
8619 pattr_hcksum_t *hck = (pattr_hcksum_t *)hck_attr.buf;
8620
8621 ASSERT(hck_attr.len >= sizeof (pattr_hcksum_t));
8622 if (flags != NULL)
8623 *flags = hck->hcksum_flags;
8624 if (start != NULL)
8625 *start = hck->hcksum_start_offset;
8626 if (stuff != NULL)
8627 *stuff = hck->hcksum_stuff_offset;
8628 if (end != NULL)
8629 *end = hck->hcksum_end_offset;
8630 if (value != NULL)
8631 *value = (uint32_t)
8632 hck->hcksum_cksum_val.inet_cksum;
8633 }
8634 }
8635 }
8636
8637 void
8638 lso_info_set(mblk_t *mp, uint32_t mss, uint32_t flags)
8639 {
8640 ASSERT(DB_TYPE(mp) == M_DATA);
8641 ASSERT((flags & ~HW_LSO_FLAGS) == 0);
8642
8643 /* Set the flags */
8644 DB_LSOFLAGS(mp) |= flags;
8645 DB_LSOMSS(mp) = mss;
8646 }
8647
8648 void
8649 lso_info_cleanup(mblk_t *mp)
8650 {
8651 ASSERT(DB_TYPE(mp) == M_DATA);
8652
8653 /* Clear the flags */
8654 DB_LSOFLAGS(mp) &= ~HW_LSO_FLAGS;
8655 DB_LSOMSS(mp) = 0;
8656 }
8657
8658 /*
8659 * Checksum buffer *bp for len bytes with psum partial checksum,
8660 * or 0 if none, and return the 16 bit partial checksum.
8661 */
8662 unsigned
8663 bcksum(uchar_t *bp, int len, unsigned int psum)
8664 {
8665 int odd = len & 1;
8666 extern unsigned int ip_ocsum();
8667
8668 if (((intptr_t)bp & 1) == 0 && !odd) {
8669 /*
8670 * Bp is 16 bit aligned and len is multiple of 16 bit word.
8671 */
8672 return (ip_ocsum((ushort_t *)bp, len >> 1, psum));
8673 }
8674 if (((intptr_t)bp & 1) != 0) {
8675 /*
8676 * Bp isn't 16 bit aligned.
8677 */
8678 unsigned int tsum;
8679
8680 #ifdef _LITTLE_ENDIAN
8681 psum += *bp;
8682 #else
8683 psum += *bp << 8;
8684 #endif
8685 len--;
8686 bp++;
8687 tsum = ip_ocsum((ushort_t *)bp, len >> 1, 0);
8688 psum += (tsum << 8) & 0xffff | (tsum >> 8);
8689 if (len & 1) {
8690 bp += len - 1;
8691 #ifdef _LITTLE_ENDIAN
8692 psum += *bp << 8;
8693 #else
8694 psum += *bp;
8695 #endif
8696 }
8697 } else {
8698 /*
8699 * Bp is 16 bit aligned.
8700 */
8701 psum = ip_ocsum((ushort_t *)bp, len >> 1, psum);
8702 if (odd) {
8703 bp += len - 1;
8704 #ifdef _LITTLE_ENDIAN
8705 psum += *bp;
8706 #else
8707 psum += *bp << 8;
8708 #endif
8709 }
8710 }
8711 /*
8712 * Normalize psum to 16 bits before returning the new partial
8713 * checksum. The max psum value before normalization is 0x3FDFE.
8714 */
8715 return ((psum >> 16) + (psum & 0xFFFF));
8716 }
8717
8718 boolean_t
8719 is_vmloaned_mblk(mblk_t *mp, multidata_t *mmd, pdesc_t *pd)
8720 {
8721 boolean_t rc;
8722
8723 ASSERT(DB_TYPE(mp) == M_DATA || DB_TYPE(mp) == M_MULTIDATA);
8724 if (DB_TYPE(mp) == M_DATA) {
8725 rc = (((mp)->b_datap->db_struioflag & STRUIO_ZC) != 0);
8726 } else {
8727 pattrinfo_t zcopy_attr = {PATTR_ZCOPY};
8728
8729 ASSERT(mmd != NULL);
8730 rc = (mmd_getpattr(mmd, pd, &zcopy_attr) != NULL);
8731 }
8732 return (rc);
8733 }
8734
8735 void
8736 freemsgchain(mblk_t *mp)
8737 {
8738 mblk_t *next;
8739
8740 while (mp != NULL) {
8741 next = mp->b_next;
8742 mp->b_next = NULL;
8743
8744 freemsg(mp);
8745 mp = next;
8746 }
8747 }
8748
8749 mblk_t *
8750 copymsgchain(mblk_t *mp)
8751 {
8752 mblk_t *nmp = NULL;
8753 mblk_t **nmpp = &nmp;
8754
8755 for (; mp != NULL; mp = mp->b_next) {
8756 if ((*nmpp = copymsg(mp)) == NULL) {
8757 freemsgchain(nmp);
8758 return (NULL);
8759 }
8760
8761 nmpp = &((*nmpp)->b_next);
8762 }
8763
8764 return (nmp);
8765 }
8766
8767 /* NOTE: Do not add code after this point. */
8768 #undef QLOCK
8769
8770 /*
8771 * Replacement for QLOCK macro for those that can't use it.
8772 */
8773 kmutex_t *
8774 QLOCK(queue_t *q)
8775 {
8776 return (&(q)->q_lock);
8777 }
8778
8779 /*
8780 * Dummy runqueues/queuerun functions functions for backwards compatibility.
8781 */
8782 #undef runqueues
8783 void
8784 runqueues(void)
8785 {
8786 }
8787
8788 #undef queuerun
8789 void
8790 queuerun(void)
8791 {
8792 }
8793
8794 /*
8795 * Initialize the STR stack instance, which tracks autopush and persistent
8796 * links.
8797 */
8798 /* ARGSUSED */
8799 static void *
8800 str_stack_init(netstackid_t stackid, netstack_t *ns)
8801 {
8802 str_stack_t *ss;
8803 int i;
8804
8805 ss = (str_stack_t *)kmem_zalloc(sizeof (*ss), KM_SLEEP);
8806 ss->ss_netstack = ns;
8807
8808 /*
8809 * set up autopush
8810 */
8811 sad_initspace(ss);
8812
8813 /*
8814 * set up mux_node structures.
8815 */
8816 ss->ss_devcnt = devcnt; /* In case it should change before free */
8817 ss->ss_mux_nodes = kmem_zalloc((sizeof (struct mux_node) *
8818 ss->ss_devcnt), KM_SLEEP);
8819 for (i = 0; i < ss->ss_devcnt; i++)
8820 ss->ss_mux_nodes[i].mn_imaj = i;
8821 return (ss);
8822 }
8823
8824 /*
8825 * Note: run at zone shutdown and not destroy so that the PLINKs are
8826 * gone by the time other cleanup happens from the destroy callbacks.
8827 */
8828 static void
8829 str_stack_shutdown(netstackid_t stackid, void *arg)
8830 {
8831 str_stack_t *ss = (str_stack_t *)arg;
8832 int i;
8833 cred_t *cr;
8834
8835 cr = zone_get_kcred(netstackid_to_zoneid(stackid));
8836 ASSERT(cr != NULL);
8837
8838 /* Undo all the I_PLINKs for this zone */
8839 for (i = 0; i < ss->ss_devcnt; i++) {
8840 struct mux_edge *ep;
8841 ldi_handle_t lh;
8842 ldi_ident_t li;
8843 int ret;
8844 int rval;
8845 dev_t rdev;
8846
8847 ep = ss->ss_mux_nodes[i].mn_outp;
8848 if (ep == NULL)
8849 continue;
8850 ret = ldi_ident_from_major((major_t)i, &li);
8851 if (ret != 0) {
8852 continue;
8853 }
8854 rdev = ep->me_dev;
8855 ret = ldi_open_by_dev(&rdev, OTYP_CHR, FREAD|FWRITE,
8856 cr, &lh, li);
8857 if (ret != 0) {
8858 ldi_ident_release(li);
8859 continue;
8860 }
8861
8862 ret = ldi_ioctl(lh, I_PUNLINK, (intptr_t)MUXID_ALL, FKIOCTL,
8863 cr, &rval);
8864 if (ret) {
8865 (void) ldi_close(lh, FREAD|FWRITE, cr);
8866 ldi_ident_release(li);
8867 continue;
8868 }
8869 (void) ldi_close(lh, FREAD|FWRITE, cr);
8870
8871 /* Close layered handles */
8872 ldi_ident_release(li);
8873 }
8874 crfree(cr);
8875
8876 sad_freespace(ss);
8877
8878 kmem_free(ss->ss_mux_nodes, sizeof (struct mux_node) * ss->ss_devcnt);
8879 ss->ss_mux_nodes = NULL;
8880 }
8881
8882 /*
8883 * Free the structure; str_stack_shutdown did the other cleanup work.
8884 */
8885 /* ARGSUSED */
8886 static void
8887 str_stack_fini(netstackid_t stackid, void *arg)
8888 {
8889 str_stack_t *ss = (str_stack_t *)arg;
8890
8891 kmem_free(ss, sizeof (*ss));
8892 }