xref: /titanic_41/usr/src/uts/common/os/ddi.c (revision d3d50737e566cade9a08d73d2af95105ac7cd960)
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 2009 Sun Microsystems, Inc.  All rights reserved.
27  * Use is subject to license terms.
28  */
29 
30 /*
31  * UNIX Device Driver Interface functions
32  *
33  * This file contains functions that are to be added to the kernel
34  * to put the interface presented to drivers in conformance with
35  * the DDI standard. Of the functions added to the kernel, 17 are
36  * function equivalents of existing macros in sysmacros.h,
37  * stream.h, and param.h
38  *
39  * 17 additional functions -- drv_getparm(), drv_setparm(),
40  * getrbuf(), freerbuf(),
41  * getemajor(), geteminor(), etoimajor(), itoemajor(), drv_usectohz(),
42  * drv_hztousec(), drv_usecwait(), drv_priv(), and kvtoppid() --
43  * are specified by DDI to exist in the kernel and are implemented here.
44  *
45  * Note that putnext() and put() are not in this file. The C version of
46  * these routines are in uts/common/os/putnext.c and assembly versions
47  * might exist for some architectures.
48  */
49 
50 #include <sys/types.h>
51 #include <sys/param.h>
52 #include <sys/t_lock.h>
53 #include <sys/time.h>
54 #include <sys/systm.h>
55 #include <sys/cpuvar.h>
56 #include <sys/signal.h>
57 #include <sys/pcb.h>
58 #include <sys/user.h>
59 #include <sys/errno.h>
60 #include <sys/buf.h>
61 #include <sys/proc.h>
62 #include <sys/cmn_err.h>
63 #include <sys/stream.h>
64 #include <sys/strsubr.h>
65 #include <sys/uio.h>
66 #include <sys/kmem.h>
67 #include <sys/conf.h>
68 #include <sys/cred.h>
69 #include <sys/vnode.h>
70 #include <sys/file.h>
71 #include <sys/poll.h>
72 #include <sys/session.h>
73 #include <sys/ddi.h>
74 #include <sys/sunddi.h>
75 #include <sys/esunddi.h>
76 #include <sys/mkdev.h>
77 #include <sys/debug.h>
78 #include <sys/vtrace.h>
79 
80 /*
81  * return internal major number corresponding to device
82  * number (new format) argument
83  */
84 major_t
getmajor(dev_t dev)85 getmajor(dev_t dev)
86 {
87 #ifdef _LP64
88 	return ((major_t)((dev >> NBITSMINOR64) & MAXMAJ64));
89 #else
90 	return ((major_t)((dev >> NBITSMINOR) & MAXMAJ));
91 #endif
92 }
93 
94 /*
95  * return external major number corresponding to device
96  * number (new format) argument
97  */
98 major_t
getemajor(dev_t dev)99 getemajor(dev_t dev)
100 {
101 #ifdef _LP64
102 	return ((major_t)((dev >> NBITSMINOR64) & MAXMAJ64));
103 #else
104 	return ((major_t)((dev >> NBITSMINOR) & MAXMAJ));
105 #endif
106 }
107 
108 /*
109  * return internal minor number corresponding to device
110  * number (new format) argument
111  */
112 minor_t
getminor(dev_t dev)113 getminor(dev_t dev)
114 {
115 #ifdef _LP64
116 	return ((minor_t)(dev & MAXMIN64));
117 #else
118 	return ((minor_t)(dev & MAXMIN));
119 #endif
120 }
121 
122 /*
123  * return external minor number corresponding to device
124  * number (new format) argument
125  */
126 minor_t
geteminor(dev_t dev)127 geteminor(dev_t dev)
128 {
129 #ifdef _LP64
130 	return ((minor_t)(dev & MAXMIN64));
131 #else
132 	return ((minor_t)(dev & MAXMIN));
133 #endif
134 }
135 
136 /*
137  * return internal major number corresponding to external
138  * major number.
139  */
140 int
etoimajor(major_t emajnum)141 etoimajor(major_t emajnum)
142 {
143 #ifdef _LP64
144 	if (emajnum >= devcnt)
145 		return (-1); /* invalid external major */
146 #else
147 	if (emajnum > MAXMAJ || emajnum >= devcnt)
148 		return (-1); /* invalid external major */
149 #endif
150 	return ((int)emajnum);
151 }
152 
153 /*
154  * return external major number corresponding to internal
155  * major number argument or -1 if no external major number
156  * can be found after lastemaj that maps to the internal
157  * major number. Pass a lastemaj val of -1 to start
158  * the search initially. (Typical use of this function is
159  * of the form:
160  *
161  *	lastemaj = -1;
162  *	while ((lastemaj = itoemajor(imag, lastemaj)) != -1)
163  *		{ process major number }
164  */
165 int
itoemajor(major_t imajnum,int lastemaj)166 itoemajor(major_t imajnum, int lastemaj)
167 {
168 	if (imajnum >= devcnt)
169 		return (-1);
170 
171 	/*
172 	 * if lastemaj == -1 then start from beginning of
173 	 * the (imaginary) MAJOR table
174 	 */
175 	if (lastemaj < -1)
176 		return (-1);
177 
178 	/*
179 	 * given that there's a 1-1 mapping of internal to external
180 	 * major numbers, searching is somewhat pointless ... let's
181 	 * just go there directly.
182 	 */
183 	if (++lastemaj < devcnt && imajnum < devcnt)
184 		return (imajnum);
185 	return (-1);
186 }
187 
188 /*
189  * encode external major and minor number arguments into a
190  * new format device number
191  */
192 dev_t
makedevice(major_t maj,minor_t minor)193 makedevice(major_t maj, minor_t minor)
194 {
195 #ifdef _LP64
196 	return (((dev_t)maj << NBITSMINOR64) | (minor & MAXMIN64));
197 #else
198 	return (((dev_t)maj << NBITSMINOR) | (minor & MAXMIN));
199 #endif
200 }
201 
202 /*
203  * cmpdev - compress new device format to old device format
204  */
205 o_dev_t
cmpdev(dev_t dev)206 cmpdev(dev_t dev)
207 {
208 	major_t major_d;
209 	minor_t minor_d;
210 
211 #ifdef _LP64
212 	major_d = dev >> NBITSMINOR64;
213 	minor_d = dev & MAXMIN64;
214 #else
215 	major_d = dev >> NBITSMINOR;
216 	minor_d = dev & MAXMIN;
217 #endif
218 	if (major_d > OMAXMAJ || minor_d > OMAXMIN)
219 		return ((o_dev_t)NODEV);
220 	return ((o_dev_t)((major_d << ONBITSMINOR) | minor_d));
221 }
222 
223 dev_t
expdev(dev_t dev)224 expdev(dev_t dev)
225 {
226 	major_t major_d;
227 	minor_t minor_d;
228 
229 	major_d = ((dev >> ONBITSMINOR) & OMAXMAJ);
230 	minor_d = (dev & OMAXMIN);
231 #ifdef _LP64
232 	return ((((dev_t)major_d << NBITSMINOR64) | minor_d));
233 #else
234 	return ((((dev_t)major_d << NBITSMINOR) | minor_d));
235 #endif
236 }
237 
238 /*
239  * return true (1) if the message type input is a data
240  * message type, 0 otherwise
241  */
242 #undef datamsg
243 int
datamsg(unsigned char db_type)244 datamsg(unsigned char db_type)
245 {
246 	return (db_type == M_DATA || db_type == M_PROTO ||
247 	    db_type == M_PCPROTO || db_type == M_DELAY);
248 }
249 
250 /*
251  * return a pointer to the other queue in the queue pair of qp
252  */
253 queue_t *
OTHERQ(queue_t * q)254 OTHERQ(queue_t *q)
255 {
256 	return (_OTHERQ(q));
257 }
258 
259 /*
260  * return a pointer to the read queue in the queue pair of qp.
261  */
262 queue_t *
RD(queue_t * q)263 RD(queue_t *q)
264 {
265 		return (_RD(q));
266 
267 }
268 
269 /*
270  * return a pointer to the write queue in the queue pair of qp.
271  */
272 int
SAMESTR(queue_t * q)273 SAMESTR(queue_t *q)
274 {
275 	return (_SAMESTR(q));
276 }
277 
278 /*
279  * return a pointer to the write queue in the queue pair of qp.
280  */
281 queue_t *
WR(queue_t * q)282 WR(queue_t *q)
283 {
284 	return (_WR(q));
285 }
286 
287 /*
288  * store value of kernel parameter associated with parm
289  */
290 int
drv_getparm(unsigned int parm,void * valuep)291 drv_getparm(unsigned int parm, void *valuep)
292 {
293 	proc_t	*p = curproc;
294 	time_t	now;
295 
296 	switch (parm) {
297 	case UPROCP:
298 		*(proc_t **)valuep = p;
299 		break;
300 	case PPGRP:
301 		mutex_enter(&p->p_lock);
302 		*(pid_t *)valuep = p->p_pgrp;
303 		mutex_exit(&p->p_lock);
304 		break;
305 	case LBOLT:
306 		*(clock_t *)valuep = ddi_get_lbolt();
307 		break;
308 	case TIME:
309 		if ((now = gethrestime_sec()) == 0) {
310 			timestruc_t ts;
311 			mutex_enter(&tod_lock);
312 			ts = tod_get();
313 			mutex_exit(&tod_lock);
314 			*(time_t *)valuep = ts.tv_sec;
315 		} else {
316 			*(time_t *)valuep = now;
317 		}
318 		break;
319 	case PPID:
320 		*(pid_t *)valuep = p->p_pid;
321 		break;
322 	case PSID:
323 		mutex_enter(&p->p_splock);
324 		*(pid_t *)valuep = p->p_sessp->s_sid;
325 		mutex_exit(&p->p_splock);
326 		break;
327 	case UCRED:
328 		*(cred_t **)valuep = CRED();
329 		break;
330 	default:
331 		return (-1);
332 	}
333 
334 	return (0);
335 }
336 
337 /*
338  * set value of kernel parameter associated with parm
339  */
340 int
drv_setparm(unsigned int parm,unsigned long value)341 drv_setparm(unsigned int parm, unsigned long value)
342 {
343 	switch (parm) {
344 	case SYSRINT:
345 		CPU_STATS_ADDQ(CPU, sys, rcvint, value);
346 		break;
347 	case SYSXINT:
348 		CPU_STATS_ADDQ(CPU, sys, xmtint, value);
349 		break;
350 	case SYSMINT:
351 		CPU_STATS_ADDQ(CPU, sys, mdmint, value);
352 		break;
353 	case SYSRAWC:
354 		CPU_STATS_ADDQ(CPU, sys, rawch, value);
355 		break;
356 	case SYSCANC:
357 		CPU_STATS_ADDQ(CPU, sys, canch, value);
358 		break;
359 	case SYSOUTC:
360 		CPU_STATS_ADDQ(CPU, sys, outch, value);
361 		break;
362 	default:
363 		return (-1);
364 	}
365 
366 	return (0);
367 }
368 
369 /*
370  * allocate space for buffer header and return pointer to it.
371  * preferred means of obtaining space for a local buf header.
372  * returns pointer to buf upon success, NULL for failure
373  */
374 struct buf *
getrbuf(int sleep)375 getrbuf(int sleep)
376 {
377 	struct buf *bp;
378 
379 	bp = kmem_alloc(sizeof (struct buf), sleep);
380 	if (bp == NULL)
381 		return (NULL);
382 	bioinit(bp);
383 
384 	return (bp);
385 }
386 
387 /*
388  * free up space allocated by getrbuf()
389  */
390 void
freerbuf(struct buf * bp)391 freerbuf(struct buf *bp)
392 {
393 	biofini(bp);
394 	kmem_free(bp, sizeof (struct buf));
395 }
396 
397 /*
398  * convert byte count input to logical page units
399  * (byte counts that are not a page-size multiple
400  * are rounded down)
401  */
402 pgcnt_t
btop(size_t numbytes)403 btop(size_t numbytes)
404 {
405 	return (numbytes >> PAGESHIFT);
406 }
407 
408 /*
409  * convert byte count input to logical page units
410  * (byte counts that are not a page-size multiple
411  * are rounded up)
412  */
413 pgcnt_t
btopr(size_t numbytes)414 btopr(size_t numbytes)
415 {
416 	return ((numbytes + PAGEOFFSET) >> PAGESHIFT);
417 }
418 
419 /*
420  * convert size in pages to bytes.
421  */
422 size_t
ptob(pgcnt_t numpages)423 ptob(pgcnt_t numpages)
424 {
425 	return (numpages << PAGESHIFT);
426 }
427 
428 #define	MAXCLOCK_T LONG_MAX
429 
430 /*
431  * Convert from system time units (hz) to microseconds.
432  *
433  * If ticks <= 0, return 0.
434  * If converting ticks to usecs would overflow, return MAXCLOCK_T.
435  * Otherwise, convert ticks to microseconds.
436  */
437 clock_t
drv_hztousec(clock_t ticks)438 drv_hztousec(clock_t ticks)
439 {
440 	if (ticks <= 0)
441 		return (0);
442 
443 	if (ticks > MAXCLOCK_T / usec_per_tick)
444 		return (MAXCLOCK_T);
445 
446 	return (TICK_TO_USEC(ticks));
447 }
448 
449 
450 /*
451  * Convert from microseconds to system time units (hz), rounded up.
452  *
453  * If ticks <= 0, return 0.
454  * Otherwise, convert microseconds to ticks, rounding up.
455  */
456 clock_t
drv_usectohz(clock_t microsecs)457 drv_usectohz(clock_t microsecs)
458 {
459 	if (microsecs <= 0)
460 		return (0);
461 
462 	return (USEC_TO_TICK_ROUNDUP(microsecs));
463 }
464 
465 #ifdef	sun
466 /*
467  * drv_usecwait implemented in each architecture's machine
468  * specific code somewhere. For sparc, it is the alternate entry
469  * to usec_delay (eventually usec_delay goes away). See
470  * sparc/os/ml/sparc_subr.s
471  */
472 #endif
473 
474 /*
475  * bcanputnext, canputnext assume called from timeout, bufcall,
476  * or esballoc free routines.  since these are driven by
477  * clock interrupts, instead of system calls the appropriate plumbing
478  * locks have not been acquired.
479  */
480 int
bcanputnext(queue_t * q,unsigned char band)481 bcanputnext(queue_t *q, unsigned char band)
482 {
483 	int	ret;
484 
485 	claimstr(q);
486 	ret = bcanput(q->q_next, band);
487 	releasestr(q);
488 	return (ret);
489 }
490 
491 int
canputnext(queue_t * q)492 canputnext(queue_t *q)
493 {
494 	queue_t	*qofsq = q;
495 	struct stdata *stp = STREAM(q);
496 	kmutex_t *sdlock;
497 
498 	TRACE_1(TR_FAC_STREAMS_FR, TR_CANPUTNEXT_IN,
499 	    "canputnext?:%p\n", q);
500 
501 	if (stp->sd_ciputctrl != NULL) {
502 		int ix = CPU->cpu_seqid & stp->sd_nciputctrl;
503 		sdlock = &stp->sd_ciputctrl[ix].ciputctrl_lock;
504 		mutex_enter(sdlock);
505 	} else
506 		mutex_enter(sdlock = &stp->sd_reflock);
507 
508 	/* get next module forward with a service queue */
509 	q = q->q_next->q_nfsrv;
510 	ASSERT(q != NULL);
511 
512 	/* this is for loopback transports, they should not do a canputnext */
513 	ASSERT(STRMATED(q->q_stream) || STREAM(q) == STREAM(qofsq));
514 
515 	if (!(q->q_flag & QFULL)) {
516 		mutex_exit(sdlock);
517 		TRACE_2(TR_FAC_STREAMS_FR, TR_CANPUTNEXT_OUT,
518 		    "canputnext:%p %d", q, 1);
519 		return (1);
520 	}
521 
522 	if (sdlock != &stp->sd_reflock) {
523 		mutex_exit(sdlock);
524 		mutex_enter(&stp->sd_reflock);
525 	}
526 
527 	/* the above is the most frequently used path */
528 	stp->sd_refcnt++;
529 	ASSERT(stp->sd_refcnt != 0);	/* Wraparound */
530 	mutex_exit(&stp->sd_reflock);
531 
532 	mutex_enter(QLOCK(q));
533 	if (q->q_flag & QFULL) {
534 		q->q_flag |= QWANTW;
535 		mutex_exit(QLOCK(q));
536 		TRACE_2(TR_FAC_STREAMS_FR, TR_CANPUTNEXT_OUT,
537 		    "canputnext:%p %d", q, 0);
538 		releasestr(qofsq);
539 
540 		return (0);
541 	}
542 	mutex_exit(QLOCK(q));
543 	TRACE_2(TR_FAC_STREAMS_FR, TR_CANPUTNEXT_OUT, "canputnext:%p %d", q, 1);
544 	releasestr(qofsq);
545 
546 	return (1);
547 }
548 
549 
550 /*
551  * Open has progressed to the point where it is safe to send/receive messages.
552  *
553  * "qprocson enables the put and service routines of the driver
554  * or module... Prior to the call to qprocson, the put and service
555  * routines of a newly pushed module or newly opened driver are
556  * disabled.  For the module, messages flow around it as if it
557  * were not present in the stream... qprocson must be called by
558  * the first open of a module or driver after allocation and
559  * initialization of any resource on which the put and service
560  * routines depend."
561  *
562  * Note that before calling qprocson a module/driver could itself cause its
563  * put or service procedures to be run by using put() or qenable().
564  */
565 void
qprocson(queue_t * q)566 qprocson(queue_t *q)
567 {
568 	ASSERT(q->q_flag & QREADR);
569 	/*
570 	 * Do not call insertq() if it is a re-open.  But if _QINSERTING
571 	 * is set, q_next will not be NULL and we need to call insertq().
572 	 */
573 	if ((q->q_next == NULL && WR(q)->q_next == NULL) ||
574 	    (q->q_flag & _QINSERTING))
575 		insertq(STREAM(q), q);
576 }
577 
578 /*
579  * Close has reached a point where it can no longer allow put/service
580  * into the queue.
581  *
582  * "qprocsoff disables the put and service routines of the driver
583  * or module... When the routines are disabled in a module, messages
584  * flow around the module as if it were not present in the stream.
585  * qprocsoff must be called by the close routine of a driver or module
586  * before deallocating any resources on which the driver/module's
587  * put and service routines depend.  qprocsoff will remove the
588  * queue's service routines from the list of service routines to be
589  * run and waits until any concurrent put or service routines are
590  * finished."
591  *
592  * Note that after calling qprocsoff a module/driver could itself cause its
593  * put procedures to be run by using put().
594  */
595 void
qprocsoff(queue_t * q)596 qprocsoff(queue_t *q)
597 {
598 	ASSERT(q->q_flag & QREADR);
599 	if (q->q_flag & QWCLOSE) {
600 		/* Called more than once */
601 		return;
602 	}
603 	disable_svc(q);
604 	removeq(q);
605 }
606 
607 /*
608  * "freezestr() freezes the state of the entire STREAM  containing
609  *  the  queue  pair  q.  A frozen STREAM blocks any thread
610  *  attempting to enter any open, close, put or service  routine
611  *  belonging  to  any  queue instance in the STREAM, and blocks
612  *  any thread currently within the STREAM if it attempts to put
613  *  messages  onto  or take messages off of any queue within the
614  *  STREAM (with the sole exception  of  the  caller).   Threads
615  *  blocked  by  this  mechanism  remain  so until the STREAM is
616  *  thawed by a call to unfreezestr().
617  *
618  * Use strblock to set SQ_FROZEN in all syncqs in the stream (prevents
619  * further entry into put, service, open, and close procedures) and
620  * grab (and hold) all the QLOCKs in the stream (to block putq, getq etc.)
621  *
622  * Note: this has to be the only code that acquires one QLOCK while holding
623  * another QLOCK (otherwise we would have locking hirarchy/ordering violations.)
624  */
625 void
freezestr(queue_t * q)626 freezestr(queue_t *q)
627 {
628 	struct stdata *stp = STREAM(q);
629 
630 	/*
631 	 * Increment refcnt to prevent q_next from changing during the strblock
632 	 * as well as while the stream is frozen.
633 	 */
634 	claimstr(RD(q));
635 
636 	strblock(q);
637 	ASSERT(stp->sd_freezer == NULL);
638 	stp->sd_freezer = curthread;
639 	for (q = stp->sd_wrq; q != NULL; q = SAMESTR(q) ? q->q_next : NULL) {
640 		mutex_enter(QLOCK(q));
641 		mutex_enter(QLOCK(RD(q)));
642 	}
643 }
644 
645 /*
646  * Undo what freezestr did.
647  * Have to drop the QLOCKs before the strunblock since strunblock will
648  * potentially call other put procedures.
649  */
650 void
unfreezestr(queue_t * q)651 unfreezestr(queue_t *q)
652 {
653 	struct stdata *stp = STREAM(q);
654 	queue_t	*q1;
655 
656 	for (q1 = stp->sd_wrq; q1 != NULL;
657 	    q1 = SAMESTR(q1) ? q1->q_next : NULL) {
658 		mutex_exit(QLOCK(q1));
659 		mutex_exit(QLOCK(RD(q1)));
660 	}
661 	ASSERT(stp->sd_freezer == curthread);
662 	stp->sd_freezer = NULL;
663 	strunblock(q);
664 	releasestr(RD(q));
665 }
666 
667 /*
668  * Used by open and close procedures to "sleep" waiting for messages to
669  * arrive. Note: can only be used in open and close procedures.
670  *
671  * Lower the gate and let in either messages on the syncq (if there are
672  * any) or put/service procedures.
673  *
674  * If the queue has an outer perimeter this will not prevent entry into this
675  * syncq (since outer_enter does not set SQ_WRITER on the syncq that gets the
676  * exclusive access to the outer perimeter.)
677  *
678  * Return 0 is the cv_wait_sig was interrupted; otherwise 1.
679  *
680  * It only makes sense to grab sq_putlocks for !SQ_CIOC sync queues because
681  * otherwise put entry points were not blocked in the first place. if this is
682  * SQ_CIOC then qwait is used to wait for service procedure to run since syncq
683  * is always SQ_CIPUT if it is SQ_CIOC.
684  *
685  * Note that SQ_EXCL is dropped and SQ_WANTEXITWAKEUP set in sq_flags
686  * atomically under sq_putlocks to make sure putnext will not miss a pending
687  * wakeup.
688  */
689 int
qwait_sig(queue_t * q)690 qwait_sig(queue_t *q)
691 {
692 	syncq_t		*sq, *outer;
693 	uint_t		flags;
694 	int		ret = 1;
695 	int		is_sq_cioc;
696 
697 	/*
698 	 * Perform the same operations as a leavesq(sq, SQ_OPENCLOSE)
699 	 * while detecting all cases where the perimeter is entered
700 	 * so that qwait_sig can return to the caller.
701 	 *
702 	 * Drain the syncq if possible. Otherwise reset SQ_EXCL and
703 	 * wait for a thread to leave the syncq.
704 	 */
705 	sq = q->q_syncq;
706 	ASSERT(sq);
707 	is_sq_cioc = (sq->sq_type & SQ_CIOC) ? 1 : 0;
708 	ASSERT(sq->sq_outer == NULL || sq->sq_outer->sq_flags & SQ_WRITER);
709 	outer = sq->sq_outer;
710 	/*
711 	 * XXX this does not work if there is only an outer perimeter.
712 	 * The semantics of qwait/qwait_sig are undefined in this case.
713 	 */
714 	if (outer)
715 		outer_exit(outer);
716 
717 	mutex_enter(SQLOCK(sq));
718 	if (is_sq_cioc == 0) {
719 		SQ_PUTLOCKS_ENTER(sq);
720 	}
721 	flags = sq->sq_flags;
722 	/*
723 	 * Drop SQ_EXCL and sq_count but hold the SQLOCK
724 	 * to prevent any undetected entry and exit into the perimeter.
725 	 */
726 	ASSERT(sq->sq_count > 0);
727 	sq->sq_count--;
728 
729 	if (is_sq_cioc == 0) {
730 		ASSERT(flags & SQ_EXCL);
731 		flags &= ~SQ_EXCL;
732 	}
733 	/*
734 	 * Unblock any thread blocked in an entersq or outer_enter.
735 	 * Note: we do not unblock a thread waiting in qwait/qwait_sig,
736 	 * since that could lead to livelock with two threads in
737 	 * qwait for the same (per module) inner perimeter.
738 	 */
739 	if (flags & SQ_WANTWAKEUP) {
740 		cv_broadcast(&sq->sq_wait);
741 		flags &= ~SQ_WANTWAKEUP;
742 	}
743 	sq->sq_flags = flags;
744 	if ((flags & SQ_QUEUED) && !(flags & SQ_STAYAWAY)) {
745 		if (is_sq_cioc == 0) {
746 			SQ_PUTLOCKS_EXIT(sq);
747 		}
748 		/* drain_syncq() drops SQLOCK */
749 		drain_syncq(sq);
750 		ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
751 		entersq(sq, SQ_OPENCLOSE);
752 		return (1);
753 	}
754 	/*
755 	 * Sleep on sq_exitwait to only be woken up when threads leave the
756 	 * put or service procedures. We can not sleep on sq_wait since an
757 	 * outer_exit in a qwait running in the same outer perimeter would
758 	 * cause a livelock "ping-pong" between two or more qwait'ers.
759 	 */
760 	do {
761 		sq->sq_flags |= SQ_WANTEXWAKEUP;
762 		if (is_sq_cioc == 0) {
763 			SQ_PUTLOCKS_EXIT(sq);
764 		}
765 		ret = cv_wait_sig(&sq->sq_exitwait, SQLOCK(sq));
766 		if (is_sq_cioc == 0) {
767 			SQ_PUTLOCKS_ENTER(sq);
768 		}
769 	} while (ret && (sq->sq_flags & SQ_WANTEXWAKEUP));
770 	if (is_sq_cioc == 0) {
771 		SQ_PUTLOCKS_EXIT(sq);
772 	}
773 	mutex_exit(SQLOCK(sq));
774 
775 	/*
776 	 * Re-enter the perimeters again
777 	 */
778 	entersq(sq, SQ_OPENCLOSE);
779 	return (ret);
780 }
781 
782 /*
783  * Used by open and close procedures to "sleep" waiting for messages to
784  * arrive. Note: can only be used in open and close procedures.
785  *
786  * Lower the gate and let in either messages on the syncq (if there are
787  * any) or put/service procedures.
788  *
789  * If the queue has an outer perimeter this will not prevent entry into this
790  * syncq (since outer_enter does not set SQ_WRITER on the syncq that gets the
791  * exclusive access to the outer perimeter.)
792  *
793  * It only makes sense to grab sq_putlocks for !SQ_CIOC sync queues because
794  * otherwise put entry points were not blocked in the first place. if this is
795  * SQ_CIOC then qwait is used to wait for service procedure to run since syncq
796  * is always SQ_CIPUT if it is SQ_CIOC.
797  *
798  * Note that SQ_EXCL is dropped and SQ_WANTEXITWAKEUP set in sq_flags
799  * atomically under sq_putlocks to make sure putnext will not miss a pending
800  * wakeup.
801  */
802 void
qwait(queue_t * q)803 qwait(queue_t *q)
804 {
805 	syncq_t		*sq, *outer;
806 	uint_t		flags;
807 	int		is_sq_cioc;
808 
809 	/*
810 	 * Perform the same operations as a leavesq(sq, SQ_OPENCLOSE)
811 	 * while detecting all cases where the perimeter is entered
812 	 * so that qwait can return to the caller.
813 	 *
814 	 * Drain the syncq if possible. Otherwise reset SQ_EXCL and
815 	 * wait for a thread to leave the syncq.
816 	 */
817 	sq = q->q_syncq;
818 	ASSERT(sq);
819 	is_sq_cioc = (sq->sq_type & SQ_CIOC) ? 1 : 0;
820 	ASSERT(sq->sq_outer == NULL || sq->sq_outer->sq_flags & SQ_WRITER);
821 	outer = sq->sq_outer;
822 	/*
823 	 * XXX this does not work if there is only an outer perimeter.
824 	 * The semantics of qwait/qwait_sig are undefined in this case.
825 	 */
826 	if (outer)
827 		outer_exit(outer);
828 
829 	mutex_enter(SQLOCK(sq));
830 	if (is_sq_cioc == 0) {
831 		SQ_PUTLOCKS_ENTER(sq);
832 	}
833 	flags = sq->sq_flags;
834 	/*
835 	 * Drop SQ_EXCL and sq_count but hold the SQLOCK
836 	 * to prevent any undetected entry and exit into the perimeter.
837 	 */
838 	ASSERT(sq->sq_count > 0);
839 	sq->sq_count--;
840 
841 	if (is_sq_cioc == 0) {
842 		ASSERT(flags & SQ_EXCL);
843 		flags &= ~SQ_EXCL;
844 	}
845 	/*
846 	 * Unblock any thread blocked in an entersq or outer_enter.
847 	 * Note: we do not unblock a thread waiting in qwait/qwait_sig,
848 	 * since that could lead to livelock with two threads in
849 	 * qwait for the same (per module) inner perimeter.
850 	 */
851 	if (flags & SQ_WANTWAKEUP) {
852 		cv_broadcast(&sq->sq_wait);
853 		flags &= ~SQ_WANTWAKEUP;
854 	}
855 	sq->sq_flags = flags;
856 	if ((flags & SQ_QUEUED) && !(flags & SQ_STAYAWAY)) {
857 		if (is_sq_cioc == 0) {
858 			SQ_PUTLOCKS_EXIT(sq);
859 		}
860 		/* drain_syncq() drops SQLOCK */
861 		drain_syncq(sq);
862 		ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
863 		entersq(sq, SQ_OPENCLOSE);
864 		return;
865 	}
866 	/*
867 	 * Sleep on sq_exitwait to only be woken up when threads leave the
868 	 * put or service procedures. We can not sleep on sq_wait since an
869 	 * outer_exit in a qwait running in the same outer perimeter would
870 	 * cause a livelock "ping-pong" between two or more qwait'ers.
871 	 */
872 	do {
873 		sq->sq_flags |= SQ_WANTEXWAKEUP;
874 		if (is_sq_cioc == 0) {
875 			SQ_PUTLOCKS_EXIT(sq);
876 		}
877 		cv_wait(&sq->sq_exitwait, SQLOCK(sq));
878 		if (is_sq_cioc == 0) {
879 			SQ_PUTLOCKS_ENTER(sq);
880 		}
881 	} while (sq->sq_flags & SQ_WANTEXWAKEUP);
882 	if (is_sq_cioc == 0) {
883 		SQ_PUTLOCKS_EXIT(sq);
884 	}
885 	mutex_exit(SQLOCK(sq));
886 
887 	/*
888 	 * Re-enter the perimeters again
889 	 */
890 	entersq(sq, SQ_OPENCLOSE);
891 }
892 
893 /*
894  * Used for the synchronous streams entrypoints when sleeping outside
895  * the perimeters. Must never be called from regular put entrypoint.
896  *
897  * There's no need to grab sq_putlocks here (which only exist for CIPUT sync
898  * queues). If it is CIPUT sync queue put entry points were not blocked in the
899  * first place by rwnext/infonext which are treated as put entrypoints for
900  * permiter syncronization purposes.
901  *
902  * Consolidation private.
903  */
904 boolean_t
qwait_rw(queue_t * q)905 qwait_rw(queue_t *q)
906 {
907 	syncq_t		*sq;
908 	ulong_t		flags;
909 	boolean_t	gotsignal = B_FALSE;
910 
911 	/*
912 	 * Perform the same operations as a leavesq(sq, SQ_PUT)
913 	 * while detecting all cases where the perimeter is entered
914 	 * so that qwait_rw can return to the caller.
915 	 *
916 	 * Drain the syncq if possible. Otherwise reset SQ_EXCL and
917 	 * wait for a thread to leave the syncq.
918 	 */
919 	sq = q->q_syncq;
920 	ASSERT(sq);
921 
922 	mutex_enter(SQLOCK(sq));
923 	flags = sq->sq_flags;
924 	/*
925 	 * Drop SQ_EXCL and sq_count but hold the SQLOCK until to prevent any
926 	 * undetected entry and exit into the perimeter.
927 	 */
928 	ASSERT(sq->sq_count > 0);
929 	sq->sq_count--;
930 	if (!(sq->sq_type & SQ_CIPUT)) {
931 		ASSERT(flags & SQ_EXCL);
932 		flags &= ~SQ_EXCL;
933 	}
934 	/*
935 	 * Unblock any thread blocked in an entersq or outer_enter.
936 	 * Note: we do not unblock a thread waiting in qwait/qwait_sig,
937 	 * since that could lead to livelock with two threads in
938 	 * qwait for the same (per module) inner perimeter.
939 	 */
940 	if (flags & SQ_WANTWAKEUP) {
941 		cv_broadcast(&sq->sq_wait);
942 		flags &= ~SQ_WANTWAKEUP;
943 	}
944 	sq->sq_flags = flags;
945 	if ((flags & SQ_QUEUED) && !(flags & SQ_STAYAWAY)) {
946 		/* drain_syncq() drops SQLOCK */
947 		drain_syncq(sq);
948 		ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
949 		entersq(sq, SQ_PUT);
950 		return (B_FALSE);
951 	}
952 	/*
953 	 * Sleep on sq_exitwait to only be woken up when threads leave the
954 	 * put or service procedures. We can not sleep on sq_wait since an
955 	 * outer_exit in a qwait running in the same outer perimeter would
956 	 * cause a livelock "ping-pong" between two or more qwait'ers.
957 	 */
958 	do {
959 		sq->sq_flags |= SQ_WANTEXWAKEUP;
960 		if (cv_wait_sig(&sq->sq_exitwait, SQLOCK(sq)) <= 0) {
961 			sq->sq_flags &= ~SQ_WANTEXWAKEUP;
962 			gotsignal = B_TRUE;
963 			break;
964 		}
965 	} while (sq->sq_flags & SQ_WANTEXWAKEUP);
966 	mutex_exit(SQLOCK(sq));
967 
968 	/*
969 	 * Re-enter the perimeters again
970 	 */
971 	entersq(sq, SQ_PUT);
972 	return (gotsignal);
973 }
974 
975 /*
976  * Asynchronously upgrade to exclusive access at either the inner or
977  * outer perimeter.
978  */
979 void
qwriter(queue_t * q,mblk_t * mp,void (* func)(),int perim)980 qwriter(queue_t *q, mblk_t *mp, void (*func)(), int perim)
981 {
982 	if (perim == PERIM_INNER)
983 		qwriter_inner(q, mp, func);
984 	else if (perim == PERIM_OUTER)
985 		qwriter_outer(q, mp, func);
986 	else
987 		panic("qwriter: wrong \"perimeter\" parameter");
988 }
989 
990 /*
991  * Schedule a synchronous streams timeout
992  */
993 timeout_id_t
qtimeout(queue_t * q,void (* func)(void *),void * arg,clock_t tim)994 qtimeout(queue_t *q, void (*func)(void *), void *arg, clock_t tim)
995 {
996 	syncq_t		*sq;
997 	callbparams_t	*cbp;
998 	timeout_id_t	tid;
999 
1000 	sq = q->q_syncq;
1001 	/*
1002 	 * you don't want the timeout firing before its params are set up
1003 	 * callbparams_alloc() acquires SQLOCK(sq)
1004 	 * qtimeout() can't fail and can't sleep, so panic if memory is not
1005 	 * available.
1006 	 */
1007 	cbp = callbparams_alloc(sq, func, arg, KM_NOSLEEP | KM_PANIC);
1008 	/*
1009 	 * the callbflags in the sq use the same flags. They get anded
1010 	 * in the callbwrapper to determine if a qun* of this callback type
1011 	 * is required. This is not a request to cancel.
1012 	 */
1013 	cbp->cbp_flags = SQ_CANCEL_TOUT;
1014 	/* check new timeout version return codes */
1015 	tid = timeout(qcallbwrapper, cbp, tim);
1016 	cbp->cbp_id = (callbparams_id_t)tid;
1017 	mutex_exit(SQLOCK(sq));
1018 	/* use local id because the cbp memory could be free by now */
1019 	return (tid);
1020 }
1021 
1022 bufcall_id_t
qbufcall(queue_t * q,size_t size,uint_t pri,void (* func)(void *),void * arg)1023 qbufcall(queue_t *q, size_t size, uint_t pri, void (*func)(void *), void *arg)
1024 {
1025 	syncq_t		*sq;
1026 	callbparams_t	*cbp;
1027 	bufcall_id_t	bid;
1028 
1029 	sq = q->q_syncq;
1030 	/*
1031 	 * you don't want the timeout firing before its params are set up
1032 	 * callbparams_alloc() acquires SQLOCK(sq) if successful.
1033 	 */
1034 	cbp = callbparams_alloc(sq, func, arg, KM_NOSLEEP);
1035 	if (cbp == NULL)
1036 		return ((bufcall_id_t)0);
1037 
1038 	/*
1039 	 * the callbflags in the sq use the same flags. They get anded
1040 	 * in the callbwrapper to determine if a qun* of this callback type
1041 	 * is required. This is not a request to cancel.
1042 	 */
1043 	cbp->cbp_flags = SQ_CANCEL_BUFCALL;
1044 	/* check new timeout version return codes */
1045 	bid = bufcall(size, pri, qcallbwrapper, cbp);
1046 	cbp->cbp_id = (callbparams_id_t)bid;
1047 	if (bid == 0) {
1048 		callbparams_free(sq, cbp);
1049 	}
1050 	mutex_exit(SQLOCK(sq));
1051 	/* use local id because the params memory could be free by now */
1052 	return (bid);
1053 }
1054 
1055 /*
1056  * cancel a timeout callback which enters the inner perimeter.
1057  * cancelling of all callback types on a given syncq is serialized.
1058  * the SQ_CALLB_BYPASSED flag indicates that the callback fn did
1059  * not execute. The quntimeout return value needs to reflect this.
1060  * As with out existing callback programming model - callbacks must
1061  * be cancelled before a close completes - so ensuring that the sq
1062  * is valid when the callback wrapper is executed.
1063  */
1064 clock_t
quntimeout(queue_t * q,timeout_id_t id)1065 quntimeout(queue_t *q, timeout_id_t id)
1066 {
1067 	syncq_t *sq = q->q_syncq;
1068 	clock_t ret;
1069 
1070 	mutex_enter(SQLOCK(sq));
1071 	/* callbacks are processed serially on each syncq */
1072 	while (sq->sq_callbflags & SQ_CALLB_CANCEL_MASK) {
1073 		sq->sq_flags |= SQ_WANTWAKEUP;
1074 		cv_wait(&sq->sq_wait, SQLOCK(sq));
1075 	}
1076 	sq->sq_cancelid = (callbparams_id_t)id;
1077 	sq->sq_callbflags = SQ_CANCEL_TOUT;
1078 	if (sq->sq_flags & SQ_WANTWAKEUP) {
1079 		cv_broadcast(&sq->sq_wait);
1080 		sq->sq_flags &= ~SQ_WANTWAKEUP;
1081 	}
1082 	mutex_exit(SQLOCK(sq));
1083 	ret = untimeout(id);
1084 	mutex_enter(SQLOCK(sq));
1085 	if (ret != -1) {
1086 		/* The wrapper was never called - need to free based on id */
1087 		callbparams_free_id(sq, (callbparams_id_t)id, SQ_CANCEL_TOUT);
1088 	}
1089 	if (sq->sq_callbflags & SQ_CALLB_BYPASSED) {
1090 		ret = 0;	/* this was how much time left */
1091 	}
1092 	sq->sq_callbflags = 0;
1093 	if (sq->sq_flags & SQ_WANTWAKEUP) {
1094 		cv_broadcast(&sq->sq_wait);
1095 		sq->sq_flags &= ~SQ_WANTWAKEUP;
1096 	}
1097 	mutex_exit(SQLOCK(sq));
1098 	return (ret);
1099 }
1100 
1101 
1102 void
qunbufcall(queue_t * q,bufcall_id_t id)1103 qunbufcall(queue_t *q, bufcall_id_t id)
1104 {
1105 	syncq_t *sq = q->q_syncq;
1106 
1107 	mutex_enter(SQLOCK(sq));
1108 	/* callbacks are processed serially on each syncq */
1109 	while (sq->sq_callbflags & SQ_CALLB_CANCEL_MASK) {
1110 		sq->sq_flags |= SQ_WANTWAKEUP;
1111 		cv_wait(&sq->sq_wait, SQLOCK(sq));
1112 	}
1113 	sq->sq_cancelid = (callbparams_id_t)id;
1114 	sq->sq_callbflags = SQ_CANCEL_BUFCALL;
1115 	if (sq->sq_flags & SQ_WANTWAKEUP) {
1116 		cv_broadcast(&sq->sq_wait);
1117 		sq->sq_flags &= ~SQ_WANTWAKEUP;
1118 	}
1119 	mutex_exit(SQLOCK(sq));
1120 	unbufcall(id);
1121 	mutex_enter(SQLOCK(sq));
1122 	/*
1123 	 * No indication from unbufcall if the callback has already run.
1124 	 * Always attempt to free it.
1125 	 */
1126 	callbparams_free_id(sq, (callbparams_id_t)id, SQ_CANCEL_BUFCALL);
1127 	sq->sq_callbflags = 0;
1128 	if (sq->sq_flags & SQ_WANTWAKEUP) {
1129 		cv_broadcast(&sq->sq_wait);
1130 		sq->sq_flags &= ~SQ_WANTWAKEUP;
1131 	}
1132 	mutex_exit(SQLOCK(sq));
1133 }
1134 
1135 /*
1136  * Associate the stream with an instance of the bottom driver.  This
1137  * function is called by APIs that establish or modify the hardware
1138  * association (ppa) of an open stream.  Two examples of such
1139  * post-open(9E) APIs are the dlpi(7p) DL_ATTACH_REQ message, and the
1140  * ndd(1M) "instance=" ioctl(2).  This interface may be called from a
1141  * stream driver's wput procedure and from within syncq perimeters,
1142  * so it can't block.
1143  *
1144  * The qassociate() "model" is that it should drive attach(9E), yet it
1145  * can't really do that because driving attach(9E) is a blocking
1146  * operation.  Instead, the qassociate() implementation has complex
1147  * dependencies on the implementation behavior of other parts of the
1148  * kernel to ensure all appropriate instances (ones that have not been
1149  * made inaccessible by DR) are attached at stream open() time, and
1150  * that they will not autodetach.  The code relies on the fact that an
1151  * open() of a stream that ends up using qassociate() always occurs on
1152  * a minor node created with CLONE_DEV.  The open() comes through
1153  * clnopen() and since clnopen() calls ddi_hold_installed_driver() we
1154  * attach all instances and mark them DN_NO_AUTODETACH (given
1155  * DN_DRIVER_HELD is maintained correctly).
1156  *
1157  * Since qassociate() can't really drive attach(9E), there are corner
1158  * cases where the compromise described above leads to qassociate()
1159  * returning failure.  This can happen when administrative functions
1160  * that cause detach(9E), such as "update_drv" or "modunload -i", are
1161  * performed on the driver between the time the stream was opened and
1162  * the time its hardware association was established.  Although this can
1163  * theoretically be an arbitrary amount of time, in practice the window
1164  * is usually quite small, since applications almost always issue their
1165  * hardware association request immediately after opening the stream,
1166  * and do not typically switch association while open.  When these
1167  * corner cases occur, and qassociate() finds the requested instance
1168  * detached, it will return failure.  This failure should be propagated
1169  * to the requesting administrative application using the appropriate
1170  * post-open(9E) API error mechanism.
1171  *
1172  * All qassociate() callers are expected to check for and gracefully handle
1173  * failure return, propagating errors back to the requesting administrative
1174  * application.
1175  */
1176 int
qassociate(queue_t * q,int instance)1177 qassociate(queue_t *q, int instance)
1178 {
1179 	vnode_t *vp;
1180 	major_t major;
1181 	dev_info_t *dip;
1182 
1183 	if (instance == -1) {
1184 		ddi_assoc_queue_with_devi(q, NULL);
1185 		return (0);
1186 	}
1187 
1188 	vp = STREAM(q)->sd_vnode;
1189 	major = getmajor(vp->v_rdev);
1190 	dip = ddi_hold_devi_by_instance(major, instance,
1191 	    E_DDI_HOLD_DEVI_NOATTACH);
1192 	if (dip == NULL)
1193 		return (-1);
1194 
1195 	ddi_assoc_queue_with_devi(q, dip);
1196 	ddi_release_devi(dip);
1197 	return (0);
1198 }
1199 
1200 /*
1201  * This routine is the SVR4MP 'replacement' for
1202  * hat_getkpfnum.  The only major difference is
1203  * the return value for illegal addresses - since
1204  * sunm_getkpfnum() and srmmu_getkpfnum() both
1205  * return '-1' for bogus mappings, we can (more or
1206  * less) return the value directly.
1207  */
1208 ppid_t
kvtoppid(caddr_t addr)1209 kvtoppid(caddr_t addr)
1210 {
1211 	return ((ppid_t)hat_getpfnum(kas.a_hat, addr));
1212 }
1213 
1214 /*
1215  * This is used to set the timeout value for cv_timed_wait() or
1216  * cv_timedwait_sig().
1217  */
1218 void
time_to_wait(clock_t * now,clock_t time)1219 time_to_wait(clock_t *now, clock_t time)
1220 {
1221 	*now = ddi_get_lbolt() + time;
1222 }
1223