xref: /titanic_41/usr/src/uts/common/io/devpoll.c (revision e4b86885570d77af552e9cf94f142f4d744fb8c8)
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 /*
22  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 #include <sys/types.h>
29 #include <sys/devops.h>
30 #include <sys/conf.h>
31 #include <sys/modctl.h>
32 #include <sys/sunddi.h>
33 #include <sys/stat.h>
34 #include <sys/poll_impl.h>
35 #include <sys/errno.h>
36 #include <sys/kmem.h>
37 #include <sys/mkdev.h>
38 #include <sys/debug.h>
39 #include <sys/file.h>
40 #include <sys/sysmacros.h>
41 #include <sys/systm.h>
42 #include <sys/bitmap.h>
43 #include <sys/devpoll.h>
44 #include <sys/rctl.h>
45 #include <sys/resource.h>
46 
47 #define	RESERVED	1
48 
49 /* local data struct */
50 static	dp_entry_t	**devpolltbl; 	/* dev poll entries */
51 static	size_t		dptblsize;
52 
53 static	kmutex_t	devpoll_lock;	/* lock protecting dev tbl */
54 int			devpoll_init;	/* is /dev/poll initialized already */
55 
56 /* device local functions */
57 
58 static int dpopen(dev_t *devp, int flag, int otyp, cred_t *credp);
59 static int dpwrite(dev_t dev, struct uio *uiop, cred_t *credp);
60 static int dpioctl(dev_t dev, int cmd, intptr_t arg, int mode, cred_t *credp,
61     int *rvalp);
62 static int dppoll(dev_t dev, short events, int anyyet, short *reventsp,
63     struct pollhead **phpp);
64 static int dpclose(dev_t dev, int flag, int otyp, cred_t *credp);
65 static dev_info_t *dpdevi;
66 
67 
68 static struct cb_ops    dp_cb_ops = {
69 	dpopen,			/* open */
70 	dpclose,		/* close */
71 	nodev,			/* strategy */
72 	nodev,			/* print */
73 	nodev,			/* dump */
74 	nodev,			/* read */
75 	dpwrite,		/* write */
76 	dpioctl,		/* ioctl */
77 	nodev,			/* devmap */
78 	nodev,			/* mmap */
79 	nodev,			/* segmap */
80 	dppoll,			/* poll */
81 	ddi_prop_op,		/* prop_op */
82 	(struct streamtab *)0,	/* streamtab */
83 	D_MP,			/* flags */
84 	CB_REV,			/* cb_ops revision */
85 	nodev,			/* aread */
86 	nodev			/* awrite */
87 };
88 
89 static int dpattach(dev_info_t *, ddi_attach_cmd_t);
90 static int dpdetach(dev_info_t *, ddi_detach_cmd_t);
91 static int dpinfo(dev_info_t *, ddi_info_cmd_t, void *, void **);
92 
93 static struct dev_ops dp_ops = {
94 	DEVO_REV,		/* devo_rev */
95 	0,			/* refcnt */
96 	dpinfo,			/* info */
97 	nulldev,		/* identify */
98 	nulldev,		/* probe */
99 	dpattach,		/* attach */
100 	dpdetach,		/* detach */
101 	nodev,			/* reset */
102 	&dp_cb_ops,		/* driver operations */
103 	(struct bus_ops *)NULL, /* bus operations */
104 	nulldev			/* power */
105 };
106 
107 
108 static struct modldrv modldrv = {
109 	&mod_driverops,		/* type of module - a driver */
110 	"/dev/poll driver",
111 	&dp_ops,
112 };
113 
114 static struct modlinkage modlinkage = {
115 	MODREV_1,
116 	(void *)&modldrv,
117 	NULL
118 };
119 
120 /*
121  * Locking Design
122  *
123  * The /dev/poll driver shares most of its code with poll sys call whose
124  * code is in common/syscall/poll.c. In poll(2) design, the pollcache
125  * structure is per lwp. An implicit assumption is made there that some
126  * portion of pollcache will never be touched by other lwps. E.g., in
127  * poll(2) design, no lwp will ever need to grow bitmap of other lwp.
128  * This assumption is not true for /dev/poll; hence the need for extra
129  * locking.
130  *
131  * To allow more parallelism, each /dev/poll file descriptor (indexed by
132  * minor number) has its own lock. Since read (dpioctl) is a much more
133  * frequent operation than write, we want to allow multiple reads on same
134  * /dev/poll fd. However, we prevent writes from being starved by giving
135  * priority to write operation. Theoretically writes can starve reads as
136  * well. But in practical sense this is not important because (1) writes
137  * happens less often than reads, and (2) write operation defines the
138  * content of poll fd a cache set. If writes happens so often that they
139  * can starve reads, that means the cached set is very unstable. It may
140  * not make sense to read an unstable cache set anyway. Therefore, the
141  * writers starving readers case is not handled in this design.
142  */
143 
144 int
145 _init()
146 {
147 	int	error;
148 
149 	dptblsize = DEVPOLLSIZE;
150 	devpolltbl = kmem_zalloc(sizeof (caddr_t) * dptblsize, KM_SLEEP);
151 	mutex_init(&devpoll_lock, NULL, MUTEX_DEFAULT, NULL);
152 	devpoll_init = 1;
153 	if ((error = mod_install(&modlinkage)) != 0) {
154 		mutex_destroy(&devpoll_lock);
155 		kmem_free(devpolltbl, sizeof (caddr_t) * dptblsize);
156 		devpoll_init = 0;
157 	}
158 	return (error);
159 }
160 
161 int
162 _fini()
163 {
164 	int error;
165 
166 	if ((error = mod_remove(&modlinkage)) != 0) {
167 		return (error);
168 	}
169 	mutex_destroy(&devpoll_lock);
170 	kmem_free(devpolltbl, sizeof (caddr_t) * dptblsize);
171 	return (0);
172 }
173 
174 int
175 _info(struct modinfo *modinfop)
176 {
177 	return (mod_info(&modlinkage, modinfop));
178 }
179 
180 /*ARGSUSED*/
181 static int
182 dpattach(dev_info_t *devi, ddi_attach_cmd_t cmd)
183 {
184 	if (ddi_create_minor_node(devi, "poll", S_IFCHR, 0, DDI_PSEUDO, NULL)
185 	    == DDI_FAILURE) {
186 		ddi_remove_minor_node(devi, NULL);
187 		return (DDI_FAILURE);
188 	}
189 	dpdevi = devi;
190 	return (DDI_SUCCESS);
191 }
192 
193 static int
194 dpdetach(dev_info_t *devi, ddi_detach_cmd_t cmd)
195 {
196 	if (cmd != DDI_DETACH)
197 		return (DDI_FAILURE);
198 
199 	ddi_remove_minor_node(devi, NULL);
200 	return (DDI_SUCCESS);
201 }
202 
203 /* ARGSUSED */
204 static int
205 dpinfo(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
206 {
207 	int error;
208 
209 	switch (infocmd) {
210 	case DDI_INFO_DEVT2DEVINFO:
211 		*result = (void *)dpdevi;
212 		error = DDI_SUCCESS;
213 		break;
214 	case DDI_INFO_DEVT2INSTANCE:
215 		*result = (void *)0;
216 		error = DDI_SUCCESS;
217 		break;
218 	default:
219 		error = DDI_FAILURE;
220 	}
221 	return (error);
222 }
223 
224 /*
225  * dp_pcache_poll has similar logic to pcache_poll() in poll.c. The major
226  * differences are: (1) /dev/poll requires scanning the bitmap starting at
227  * where it was stopped last time, instead of always starting from 0,
228  * (2) since user may not have cleaned up the cached fds when they are
229  * closed, some polldats in cache may refer to closed or reused fds. We
230  * need to check for those cases.
231  *
232  * NOTE: Upon closing an fd, automatic poll cache cleanup is done for
233  *	 poll(2) caches but NOT for /dev/poll caches. So expect some
234  *	 stale entries!
235  */
236 static int
237 dp_pcache_poll(pollfd_t *pfdp, pollcache_t *pcp, nfds_t nfds, int *fdcntp)
238 {
239 	int		start, ostart, end;
240 	int		fdcnt, fd;
241 	boolean_t 	done;
242 	file_t		*fp;
243 	short		revent;
244 	boolean_t	no_wrap;
245 	pollhead_t	*php;
246 	polldat_t	*pdp;
247 	int		error = 0;
248 
249 	ASSERT(MUTEX_HELD(&pcp->pc_lock));
250 	if (pcp->pc_bitmap == NULL) {
251 		/*
252 		 * No Need to search because no poll fd
253 		 * has been cached.
254 		 */
255 		return (error);
256 	}
257 retry:
258 	start = ostart = pcp->pc_mapstart;
259 	end = pcp->pc_mapend;
260 	php = NULL;
261 
262 	if (start == 0) {
263 		/*
264 		 * started from every begining, no need to wrap around.
265 		 */
266 		no_wrap = B_TRUE;
267 	} else {
268 		no_wrap = B_FALSE;
269 	}
270 	done = B_FALSE;
271 	fdcnt = 0;
272 	while ((fdcnt < nfds) && !done) {
273 		php = NULL;
274 		revent = 0;
275 		/*
276 		 * Examine the bit map in a circular fashion
277 		 * to avoid starvation. Always resume from
278 		 * last stop. Scan till end of the map. Then
279 		 * wrap around.
280 		 */
281 		fd = bt_getlowbit(pcp->pc_bitmap, start, end);
282 		ASSERT(fd <= end);
283 		if (fd >= 0) {
284 			if (fd == end) {
285 				if (no_wrap) {
286 					done = B_TRUE;
287 				} else {
288 					start = 0;
289 					end = ostart - 1;
290 					no_wrap = B_TRUE;
291 				}
292 			} else {
293 				start = fd + 1;
294 			}
295 			pdp = pcache_lookup_fd(pcp, fd);
296 repoll:
297 			ASSERT(pdp != NULL);
298 			ASSERT(pdp->pd_fd == fd);
299 			if (pdp->pd_fp == NULL) {
300 				/*
301 				 * The fd is POLLREMOVed. This fd is
302 				 * logically no longer cached. So move
303 				 * on to the next one.
304 				 */
305 				continue;
306 			}
307 			if ((fp = getf(fd)) == NULL) {
308 				/*
309 				 * The fd has been closed, but user has not
310 				 * done a POLLREMOVE on this fd yet. Instead
311 				 * of cleaning it here implicitly, we return
312 				 * POLLNVAL. This is consistent with poll(2)
313 				 * polling a closed fd. Hope this will remind
314 				 * user to do a POLLREMOVE.
315 				 */
316 				pfdp[fdcnt].fd = fd;
317 				pfdp[fdcnt].revents = POLLNVAL;
318 				fdcnt++;
319 				continue;
320 			}
321 			if (fp != pdp->pd_fp) {
322 				/*
323 				 * user is polling on a cached fd which was
324 				 * closed and then reused. Unfortunately
325 				 * there is no good way to inform user.
326 				 * If the file struct is also reused, we
327 				 * may not be able to detect the fd reuse
328 				 * at all.  As long as this does not
329 				 * cause system failure and/or memory leak,
330 				 * we will play along. Man page states if
331 				 * user does not clean up closed fds, polling
332 				 * results will be indeterministic.
333 				 *
334 				 * XXX - perhaps log the detection of fd
335 				 *	 reuse?
336 				 */
337 				pdp->pd_fp = fp;
338 			}
339 			/*
340 			 * XXX - pollrelock() logic needs to know which
341 			 * which pollcache lock to grab. It'd be a
342 			 * cleaner solution if we could pass pcp as
343 			 * an arguement in VOP_POLL interface instead
344 			 * of implicitly passing it using thread_t
345 			 * struct. On the other hand, changing VOP_POLL
346 			 * interface will require all driver/file system
347 			 * poll routine to change. May want to revisit
348 			 * the tradeoff later.
349 			 */
350 			curthread->t_pollcache = pcp;
351 			error = VOP_POLL(fp->f_vnode, pdp->pd_events, 0,
352 			    &revent, &php, NULL);
353 			curthread->t_pollcache = NULL;
354 			releasef(fd);
355 			if (error != 0) {
356 				break;
357 			}
358 			/*
359 			 * layered devices (e.g. console driver)
360 			 * may change the vnode and thus the pollhead
361 			 * pointer out from underneath us.
362 			 */
363 			if (php != NULL && pdp->pd_php != NULL &&
364 			    php != pdp->pd_php) {
365 				pollhead_delete(pdp->pd_php, pdp);
366 				pdp->pd_php = php;
367 				pollhead_insert(php, pdp);
368 				/*
369 				 * The bit should still be set.
370 				 */
371 				ASSERT(BT_TEST(pcp->pc_bitmap, fd));
372 				goto retry;
373 			}
374 
375 			if (revent != 0) {
376 				pfdp[fdcnt].fd = fd;
377 				pfdp[fdcnt].events = pdp->pd_events;
378 				pfdp[fdcnt].revents = revent;
379 				fdcnt++;
380 			} else if (php != NULL) {
381 				/*
382 				 * We clear a bit or cache a poll fd if
383 				 * the driver returns a poll head ptr,
384 				 * which is expected in the case of 0
385 				 * revents. Some buggy driver may return
386 				 * NULL php pointer with 0 revents. In
387 				 * this case, we just treat the driver as
388 				 * "noncachable" and not clearing the bit
389 				 * in bitmap.
390 				 */
391 				if ((pdp->pd_php != NULL) &&
392 				    ((pcp->pc_flag & T_POLLWAKE) == 0)) {
393 					BT_CLEAR(pcp->pc_bitmap, fd);
394 				}
395 				if (pdp->pd_php == NULL) {
396 					pollhead_insert(php, pdp);
397 					pdp->pd_php = php;
398 					/*
399 					 * An event of interest may have
400 					 * arrived between the VOP_POLL() and
401 					 * the pollhead_insert(); check again.
402 					 */
403 					goto repoll;
404 				}
405 			}
406 		} else {
407 			/*
408 			 * No bit set in the range. Check for wrap around.
409 			 */
410 			if (!no_wrap) {
411 				start = 0;
412 				end = ostart - 1;
413 				no_wrap = B_TRUE;
414 			} else {
415 				done = B_TRUE;
416 			}
417 		}
418 	}
419 
420 	if (!done) {
421 		pcp->pc_mapstart = start;
422 	}
423 	ASSERT(*fdcntp == 0);
424 	*fdcntp = fdcnt;
425 	return (error);
426 }
427 
428 /*ARGSUSED*/
429 static int
430 dpopen(dev_t *devp, int flag, int otyp, cred_t *credp)
431 {
432 	minor_t		minordev;
433 	dp_entry_t	*dpep;
434 	pollcache_t	*pcp;
435 
436 	ASSERT(devpoll_init);
437 	ASSERT(dptblsize <= MAXMIN);
438 	mutex_enter(&devpoll_lock);
439 	for (minordev = 0; minordev < dptblsize; minordev++) {
440 		if (devpolltbl[minordev] == NULL) {
441 			devpolltbl[minordev] = (dp_entry_t *)RESERVED;
442 			break;
443 		}
444 	}
445 	if (minordev == dptblsize) {
446 		dp_entry_t	**newtbl;
447 		size_t		oldsize;
448 
449 		/*
450 		 * Used up every entry in the existing devpoll table.
451 		 * Grow the table by DEVPOLLSIZE.
452 		 */
453 		if ((oldsize = dptblsize) >= MAXMIN) {
454 			mutex_exit(&devpoll_lock);
455 			return (ENXIO);
456 		}
457 		dptblsize += DEVPOLLSIZE;
458 		if (dptblsize > MAXMIN) {
459 			dptblsize = MAXMIN;
460 		}
461 		newtbl = kmem_zalloc(sizeof (caddr_t) * dptblsize, KM_SLEEP);
462 		bcopy(devpolltbl, newtbl, sizeof (caddr_t) * oldsize);
463 		kmem_free(devpolltbl, sizeof (caddr_t) * oldsize);
464 		devpolltbl = newtbl;
465 		devpolltbl[minordev] = (dp_entry_t *)RESERVED;
466 	}
467 	mutex_exit(&devpoll_lock);
468 
469 	dpep = kmem_zalloc(sizeof (dp_entry_t), KM_SLEEP);
470 	/*
471 	 * allocate a pollcache skeleton here. Delay allocating bitmap
472 	 * structures until dpwrite() time, since we don't know the
473 	 * optimal size yet.
474 	 */
475 	pcp = pcache_alloc();
476 	dpep->dpe_pcache = pcp;
477 	pcp->pc_pid = curproc->p_pid;
478 	*devp = makedevice(getmajor(*devp), minordev);  /* clone the driver */
479 	mutex_enter(&devpoll_lock);
480 	ASSERT(minordev < dptblsize);
481 	ASSERT(devpolltbl[minordev] == (dp_entry_t *)RESERVED);
482 	devpolltbl[minordev] = dpep;
483 	mutex_exit(&devpoll_lock);
484 	return (0);
485 }
486 
487 /*
488  * Write to dev/poll add/remove fd's to/from a cached poll fd set,
489  * or change poll events for a watched fd.
490  */
491 /*ARGSUSED*/
492 static int
493 dpwrite(dev_t dev, struct uio *uiop, cred_t *credp)
494 {
495 	minor_t 	minor;
496 	dp_entry_t	*dpep;
497 	pollcache_t	*pcp;
498 	pollfd_t	*pollfdp, *pfdp;
499 	int		error;
500 	ssize_t		uiosize;
501 	nfds_t		pollfdnum;
502 	struct pollhead	*php = NULL;
503 	polldat_t	*pdp;
504 	int		fd;
505 	file_t		*fp;
506 
507 	minor = getminor(dev);
508 
509 	mutex_enter(&devpoll_lock);
510 	ASSERT(minor < dptblsize);
511 	dpep = devpolltbl[minor];
512 	ASSERT(dpep != NULL);
513 	mutex_exit(&devpoll_lock);
514 	pcp = dpep->dpe_pcache;
515 	if (curproc->p_pid != pcp->pc_pid) {
516 		return (EACCES);
517 	}
518 	uiosize = uiop->uio_resid;
519 	pollfdnum = uiosize / sizeof (pollfd_t);
520 	mutex_enter(&curproc->p_lock);
521 	if (pollfdnum > (uint_t)rctl_enforced_value(
522 	    rctlproc_legacy[RLIMIT_NOFILE], curproc->p_rctls, curproc)) {
523 		(void) rctl_action(rctlproc_legacy[RLIMIT_NOFILE],
524 		    curproc->p_rctls, curproc, RCA_SAFE);
525 		mutex_exit(&curproc->p_lock);
526 		return (set_errno(EINVAL));
527 	}
528 	mutex_exit(&curproc->p_lock);
529 	/*
530 	 * Copy in the pollfd array.  Walk through the array and add
531 	 * each polled fd to the cached set.
532 	 */
533 	pollfdp = kmem_alloc(uiosize, KM_SLEEP);
534 
535 	/*
536 	 * Although /dev/poll uses the write(2) interface to cache fds, it's
537 	 * not supposed to function as a seekable device. To prevent offset
538 	 * from growing and eventually exceed the maximum, reset the offset
539 	 * here for every call.
540 	 */
541 	uiop->uio_loffset = 0;
542 	if ((error = uiomove((caddr_t)pollfdp, uiosize, UIO_WRITE, uiop))
543 	    != 0) {
544 		kmem_free(pollfdp, uiosize);
545 		return (error);
546 	}
547 	/*
548 	 * We are about to enter the core portion of dpwrite(). Make sure this
549 	 * write has exclusive access in this portion of the code, i.e., no
550 	 * other writers in this code and no other readers in dpioctl.
551 	 */
552 	mutex_enter(&dpep->dpe_lock);
553 	dpep->dpe_writerwait++;
554 	while (dpep->dpe_refcnt != 0) {
555 		if (!cv_wait_sig_swap(&dpep->dpe_cv, &dpep->dpe_lock)) {
556 			dpep->dpe_writerwait--;
557 			mutex_exit(&dpep->dpe_lock);
558 			kmem_free(pollfdp, uiosize);
559 			return (set_errno(EINTR));
560 		}
561 	}
562 	dpep->dpe_writerwait--;
563 	dpep->dpe_flag |= DP_WRITER_PRESENT;
564 	dpep->dpe_refcnt++;
565 	mutex_exit(&dpep->dpe_lock);
566 
567 	mutex_enter(&pcp->pc_lock);
568 	if (pcp->pc_bitmap == NULL) {
569 		pcache_create(pcp, pollfdnum);
570 	}
571 	for (pfdp = pollfdp; pfdp < pollfdp + pollfdnum; pfdp++) {
572 		fd = pfdp->fd;
573 		if ((uint_t)fd >= P_FINFO(curproc)->fi_nfiles)
574 			continue;
575 		pdp = pcache_lookup_fd(pcp, fd);
576 		if (pfdp->events != POLLREMOVE) {
577 			if (pdp == NULL) {
578 				pdp = pcache_alloc_fd(0);
579 				pdp->pd_fd = fd;
580 				pdp->pd_pcache = pcp;
581 				pcache_insert_fd(pcp, pdp, pollfdnum);
582 			}
583 			ASSERT(pdp->pd_fd == fd);
584 			ASSERT(pdp->pd_pcache == pcp);
585 			if (fd >= pcp->pc_mapsize) {
586 				mutex_exit(&pcp->pc_lock);
587 				pcache_grow_map(pcp, fd);
588 				mutex_enter(&pcp->pc_lock);
589 			}
590 			if (fd > pcp->pc_mapend) {
591 				pcp->pc_mapend = fd;
592 			}
593 			if ((fp = getf(fd)) == NULL) {
594 				/*
595 				 * The fd is not valid. Since we can't pass
596 				 * this error back in the write() call, set
597 				 * the bit in bitmap to force DP_POLL ioctl
598 				 * to examine it.
599 				 */
600 				BT_SET(pcp->pc_bitmap, fd);
601 				pdp->pd_events |= pfdp->events;
602 				continue;
603 			}
604 			/*
605 			 * Don't do VOP_POLL for an already cached fd with
606 			 * same poll events.
607 			 */
608 			if ((pdp->pd_events == pfdp->events) &&
609 			    (pdp->pd_fp != NULL)) {
610 				/*
611 				 * the events are already cached
612 				 */
613 				releasef(fd);
614 				continue;
615 			}
616 
617 			/*
618 			 * do VOP_POLL and cache this poll fd.
619 			 */
620 			/*
621 			 * XXX - pollrelock() logic needs to know which
622 			 * which pollcache lock to grab. It'd be a
623 			 * cleaner solution if we could pass pcp as
624 			 * an arguement in VOP_POLL interface instead
625 			 * of implicitly passing it using thread_t
626 			 * struct. On the other hand, changing VOP_POLL
627 			 * interface will require all driver/file system
628 			 * poll routine to change. May want to revisit
629 			 * the tradeoff later.
630 			 */
631 			curthread->t_pollcache = pcp;
632 			error = VOP_POLL(fp->f_vnode, pfdp->events, 0,
633 			    &pfdp->revents, &php, NULL);
634 			curthread->t_pollcache = NULL;
635 			/*
636 			 * We always set the bit when this fd is cached;
637 			 * this forces the first DP_POLL to poll this fd.
638 			 * Real performance gain comes from subsequent
639 			 * DP_POLL.  We also attempt a pollhead_insert();
640 			 * if it's not possible, we'll do it in dpioctl().
641 			 */
642 			BT_SET(pcp->pc_bitmap, fd);
643 			if (error != 0) {
644 				releasef(fd);
645 				break;
646 			}
647 			pdp->pd_fp = fp;
648 			pdp->pd_events |= pfdp->events;
649 			if (php != NULL) {
650 				if (pdp->pd_php == NULL) {
651 					pollhead_insert(php, pdp);
652 					pdp->pd_php = php;
653 				} else {
654 					if (pdp->pd_php != php) {
655 						pollhead_delete(pdp->pd_php,
656 						    pdp);
657 						pollhead_insert(php, pdp);
658 						pdp->pd_php = php;
659 					}
660 				}
661 
662 			}
663 			releasef(fd);
664 		} else {
665 			if (pdp == NULL) {
666 				continue;
667 			}
668 			ASSERT(pdp->pd_fd == fd);
669 			pdp->pd_fp = NULL;
670 			pdp->pd_events = 0;
671 			ASSERT(pdp->pd_thread == NULL);
672 			if (pdp->pd_php != NULL) {
673 				pollhead_delete(pdp->pd_php, pdp);
674 				pdp->pd_php = NULL;
675 			}
676 			BT_CLEAR(pcp->pc_bitmap, fd);
677 		}
678 	}
679 	mutex_exit(&pcp->pc_lock);
680 	mutex_enter(&dpep->dpe_lock);
681 	dpep->dpe_flag &= ~DP_WRITER_PRESENT;
682 	ASSERT(dpep->dpe_refcnt == 1);
683 	dpep->dpe_refcnt--;
684 	cv_broadcast(&dpep->dpe_cv);
685 	mutex_exit(&dpep->dpe_lock);
686 	kmem_free(pollfdp, uiosize);
687 	return (error);
688 }
689 
690 /*ARGSUSED*/
691 static int
692 dpioctl(dev_t dev, int cmd, intptr_t arg, int mode, cred_t *credp, int *rvalp)
693 {
694 	timestruc_t	now;
695 	timestruc_t	rqtime;
696 	timestruc_t	*rqtp = NULL;
697 	int		timecheck = 0;
698 	minor_t 	minor;
699 	dp_entry_t	*dpep;
700 	pollcache_t	*pcp;
701 	int 		error = 0;
702 	STRUCT_DECL(dvpoll, dvpoll);
703 
704 	if (cmd == DP_POLL) {
705 		/* do this now, before we sleep on DP_WRITER_PRESENT below */
706 		timecheck = timechanged;
707 		gethrestime(&now);
708 	}
709 	minor = getminor(dev);
710 	mutex_enter(&devpoll_lock);
711 	ASSERT(minor < dptblsize);
712 	dpep = devpolltbl[minor];
713 	mutex_exit(&devpoll_lock);
714 	ASSERT(dpep != NULL);
715 	pcp = dpep->dpe_pcache;
716 	if (curproc->p_pid != pcp->pc_pid)
717 		return (EACCES);
718 
719 	mutex_enter(&dpep->dpe_lock);
720 	while ((dpep->dpe_flag & DP_WRITER_PRESENT) ||
721 	    (dpep->dpe_writerwait != 0)) {
722 		if (!cv_wait_sig_swap(&dpep->dpe_cv, &dpep->dpe_lock)) {
723 			mutex_exit(&dpep->dpe_lock);
724 			return (EINTR);
725 		}
726 	}
727 	dpep->dpe_refcnt++;
728 	mutex_exit(&dpep->dpe_lock);
729 
730 	switch (cmd) {
731 	case	DP_POLL:
732 	{
733 		pollstate_t *ps;
734 		nfds_t	nfds;
735 		int	fdcnt = 0;
736 		int	time_out;
737 		int	rval;
738 
739 		STRUCT_INIT(dvpoll, mode);
740 		error = copyin((caddr_t)arg, STRUCT_BUF(dvpoll),
741 		    STRUCT_SIZE(dvpoll));
742 		if (error) {
743 			DP_REFRELE(dpep);
744 			return (EFAULT);
745 		}
746 
747 		time_out = STRUCT_FGET(dvpoll, dp_timeout);
748 		if (time_out > 0) {
749 			/*
750 			 * Determine the future time of the requested timeout.
751 			 */
752 			rqtp = &rqtime;
753 			rqtp->tv_sec = time_out / MILLISEC;
754 			rqtp->tv_nsec = (time_out % MILLISEC) * MICROSEC;
755 			timespecadd(rqtp, &now);
756 		}
757 
758 		if ((nfds = STRUCT_FGET(dvpoll, dp_nfds)) == 0) {
759 			/*
760 			 * We are just using DP_POLL to sleep, so
761 			 * we don't any of the devpoll apparatus.
762 			 * Do not check for signals if we have a zero timeout.
763 			 */
764 			DP_REFRELE(dpep);
765 			if (time_out == 0)
766 				return (0);
767 			mutex_enter(&curthread->t_delay_lock);
768 			while ((rval = cv_waituntil_sig(&curthread->t_delay_cv,
769 			    &curthread->t_delay_lock, rqtp, timecheck)) > 0)
770 				continue;
771 			mutex_exit(&curthread->t_delay_lock);
772 			return ((rval == 0)? EINTR : 0);
773 		}
774 
775 		/*
776 		 * XXX It'd be nice not to have to alloc each time.
777 		 * But it requires another per thread structure hook.
778 		 * Do it later if there is data suggest that.
779 		 */
780 		if ((ps = curthread->t_pollstate) == NULL) {
781 			curthread->t_pollstate = pollstate_create();
782 			ps = curthread->t_pollstate;
783 		}
784 		if (ps->ps_dpbufsize < nfds) {
785 			struct proc *p = ttoproc(curthread);
786 			/*
787 			 * The maximum size should be no large than
788 			 * current maximum open file count.
789 			 */
790 			mutex_enter(&p->p_lock);
791 			if (nfds > p->p_fno_ctl) {
792 				mutex_exit(&p->p_lock);
793 				DP_REFRELE(dpep);
794 				return (EINVAL);
795 			}
796 			mutex_exit(&p->p_lock);
797 			kmem_free(ps->ps_dpbuf, sizeof (pollfd_t) *
798 			    ps->ps_dpbufsize);
799 			ps->ps_dpbuf = kmem_zalloc(sizeof (pollfd_t) *
800 			    nfds, KM_SLEEP);
801 			ps->ps_dpbufsize = nfds;
802 		}
803 
804 		mutex_enter(&pcp->pc_lock);
805 		for (;;) {
806 			pcp->pc_flag = 0;
807 			error = dp_pcache_poll(ps->ps_dpbuf, pcp, nfds, &fdcnt);
808 			if (fdcnt > 0 || error != 0)
809 				break;
810 
811 			/*
812 			 * A pollwake has happened since we polled cache.
813 			 */
814 			if (pcp->pc_flag & T_POLLWAKE)
815 				continue;
816 
817 			/*
818 			 * Sleep until we are notified, signaled, or timed out.
819 			 * Do not check for signals if we have a zero timeout.
820 			 */
821 			if (time_out == 0)	/* immediate timeout */
822 				break;
823 			rval = cv_waituntil_sig(&pcp->pc_cv, &pcp->pc_lock,
824 			    rqtp, timecheck);
825 			/*
826 			 * If we were awakened by a signal or timeout
827 			 * then break the loop, else poll again.
828 			 */
829 			if (rval <= 0) {
830 				if (rval == 0)	/* signal */
831 					error = EINTR;
832 				break;
833 			}
834 		}
835 		mutex_exit(&pcp->pc_lock);
836 
837 		if (error == 0 && fdcnt > 0) {
838 			if (copyout(ps->ps_dpbuf, STRUCT_FGETP(dvpoll,
839 			    dp_fds), sizeof (pollfd_t) * fdcnt)) {
840 				DP_REFRELE(dpep);
841 				return (EFAULT);
842 			}
843 			*rvalp = fdcnt;
844 		}
845 		break;
846 	}
847 
848 	case	DP_ISPOLLED:
849 	{
850 		pollfd_t	pollfd;
851 		polldat_t	*pdp;
852 
853 		STRUCT_INIT(dvpoll, mode);
854 		error = copyin((caddr_t)arg, &pollfd, sizeof (pollfd_t));
855 		if (error) {
856 			DP_REFRELE(dpep);
857 			return (EFAULT);
858 		}
859 		mutex_enter(&pcp->pc_lock);
860 		if (pcp->pc_hash == NULL) {
861 			/*
862 			 * No Need to search because no poll fd
863 			 * has been cached.
864 			 */
865 			mutex_exit(&pcp->pc_lock);
866 			DP_REFRELE(dpep);
867 			return (0);
868 		}
869 		if (pollfd.fd < 0) {
870 			mutex_exit(&pcp->pc_lock);
871 			break;
872 		}
873 		pdp = pcache_lookup_fd(pcp, pollfd.fd);
874 		if ((pdp != NULL) && (pdp->pd_fd == pollfd.fd) &&
875 		    (pdp->pd_fp != NULL)) {
876 			pollfd.revents = pdp->pd_events;
877 			if (copyout(&pollfd, (caddr_t)arg, sizeof (pollfd_t))) {
878 				mutex_exit(&pcp->pc_lock);
879 				DP_REFRELE(dpep);
880 				return (EFAULT);
881 			}
882 			*rvalp = 1;
883 		}
884 		mutex_exit(&pcp->pc_lock);
885 		break;
886 	}
887 
888 	default:
889 		DP_REFRELE(dpep);
890 		return (EINVAL);
891 	}
892 	DP_REFRELE(dpep);
893 	return (error);
894 }
895 
896 /*ARGSUSED*/
897 static int
898 dppoll(dev_t dev, short events, int anyyet, short *reventsp,
899     struct pollhead **phpp)
900 {
901 	/*
902 	 * Polling on a /dev/poll fd is not fully supported yet.
903 	 */
904 	*reventsp = POLLERR;
905 	return (0);
906 }
907 
908 /*
909  * devpoll close should do enough clean up before the pollcache is deleted,
910  * i.e., it should ensure no one still references the pollcache later.
911  * There is no "permission" check in here. Any process having the last
912  * reference of this /dev/poll fd can close.
913  */
914 /*ARGSUSED*/
915 static int
916 dpclose(dev_t dev, int flag, int otyp, cred_t *credp)
917 {
918 	minor_t 	minor;
919 	dp_entry_t	*dpep;
920 	pollcache_t	*pcp;
921 	int		i;
922 	polldat_t	**hashtbl;
923 	polldat_t	*pdp;
924 
925 	minor = getminor(dev);
926 
927 	mutex_enter(&devpoll_lock);
928 	dpep = devpolltbl[minor];
929 	ASSERT(dpep != NULL);
930 	devpolltbl[minor] = NULL;
931 	mutex_exit(&devpoll_lock);
932 	pcp = dpep->dpe_pcache;
933 	ASSERT(pcp != NULL);
934 	/*
935 	 * At this point, no other lwp can access this pollcache via the
936 	 * /dev/poll fd. This pollcache is going away, so do the clean
937 	 * up without the pc_lock.
938 	 */
939 	hashtbl = pcp->pc_hash;
940 	for (i = 0; i < pcp->pc_hashsize; i++) {
941 		for (pdp = hashtbl[i]; pdp; pdp = pdp->pd_hashnext) {
942 			if (pdp->pd_php != NULL) {
943 				pollhead_delete(pdp->pd_php, pdp);
944 				pdp->pd_php = NULL;
945 				pdp->pd_fp = NULL;
946 			}
947 		}
948 	}
949 	/*
950 	 * pollwakeup() may still interact with this pollcache. Wait until
951 	 * it is done.
952 	 */
953 	mutex_enter(&pcp->pc_no_exit);
954 	ASSERT(pcp->pc_busy >= 0);
955 	while (pcp->pc_busy > 0)
956 		cv_wait(&pcp->pc_busy_cv, &pcp->pc_no_exit);
957 	mutex_exit(&pcp->pc_no_exit);
958 	pcache_destroy(pcp);
959 	ASSERT(dpep->dpe_refcnt == 0);
960 	kmem_free(dpep, sizeof (dp_entry_t));
961 	return (0);
962 }
963