xref: /titanic_52/usr/src/uts/common/io/devpoll.c (revision ebb8ac078e9265f87093fbb363e8c2cbc6ee13e6)
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 /*
27  * Copyright (c) 2012 by Delphix. All rights reserved.
28  */
29 
30 #include <sys/types.h>
31 #include <sys/devops.h>
32 #include <sys/conf.h>
33 #include <sys/modctl.h>
34 #include <sys/sunddi.h>
35 #include <sys/stat.h>
36 #include <sys/poll_impl.h>
37 #include <sys/errno.h>
38 #include <sys/kmem.h>
39 #include <sys/mkdev.h>
40 #include <sys/debug.h>
41 #include <sys/file.h>
42 #include <sys/sysmacros.h>
43 #include <sys/systm.h>
44 #include <sys/bitmap.h>
45 #include <sys/devpoll.h>
46 #include <sys/rctl.h>
47 #include <sys/resource.h>
48 
49 #define	RESERVED	1
50 
51 /* local data struct */
52 static	dp_entry_t	**devpolltbl;	/* dev poll entries */
53 static	size_t		dptblsize;
54 
55 static	kmutex_t	devpoll_lock;	/* lock protecting dev tbl */
56 int			devpoll_init;	/* is /dev/poll initialized already */
57 
58 /* device local functions */
59 
60 static int dpopen(dev_t *devp, int flag, int otyp, cred_t *credp);
61 static int dpwrite(dev_t dev, struct uio *uiop, cred_t *credp);
62 static int dpioctl(dev_t dev, int cmd, intptr_t arg, int mode, cred_t *credp,
63     int *rvalp);
64 static int dppoll(dev_t dev, short events, int anyyet, short *reventsp,
65     struct pollhead **phpp);
66 static int dpclose(dev_t dev, int flag, int otyp, cred_t *credp);
67 static dev_info_t *dpdevi;
68 
69 
70 static struct cb_ops    dp_cb_ops = {
71 	dpopen,			/* open */
72 	dpclose,		/* close */
73 	nodev,			/* strategy */
74 	nodev,			/* print */
75 	nodev,			/* dump */
76 	nodev,			/* read */
77 	dpwrite,		/* write */
78 	dpioctl,		/* ioctl */
79 	nodev,			/* devmap */
80 	nodev,			/* mmap */
81 	nodev,			/* segmap */
82 	dppoll,			/* poll */
83 	ddi_prop_op,		/* prop_op */
84 	(struct streamtab *)0,	/* streamtab */
85 	D_MP,			/* flags */
86 	CB_REV,			/* cb_ops revision */
87 	nodev,			/* aread */
88 	nodev			/* awrite */
89 };
90 
91 static int dpattach(dev_info_t *, ddi_attach_cmd_t);
92 static int dpdetach(dev_info_t *, ddi_detach_cmd_t);
93 static int dpinfo(dev_info_t *, ddi_info_cmd_t, void *, void **);
94 
95 static struct dev_ops dp_ops = {
96 	DEVO_REV,		/* devo_rev */
97 	0,			/* refcnt */
98 	dpinfo,			/* info */
99 	nulldev,		/* identify */
100 	nulldev,		/* probe */
101 	dpattach,		/* attach */
102 	dpdetach,		/* detach */
103 	nodev,			/* reset */
104 	&dp_cb_ops,		/* driver operations */
105 	(struct bus_ops *)NULL, /* bus operations */
106 	nulldev,		/* power */
107 	ddi_quiesce_not_needed,		/* quiesce */
108 };
109 
110 
111 static struct modldrv modldrv = {
112 	&mod_driverops,		/* type of module - a driver */
113 	"/dev/poll driver",
114 	&dp_ops,
115 };
116 
117 static struct modlinkage modlinkage = {
118 	MODREV_1,
119 	(void *)&modldrv,
120 	NULL
121 };
122 
123 /*
124  * Locking Design
125  *
126  * The /dev/poll driver shares most of its code with poll sys call whose
127  * code is in common/syscall/poll.c. In poll(2) design, the pollcache
128  * structure is per lwp. An implicit assumption is made there that some
129  * portion of pollcache will never be touched by other lwps. E.g., in
130  * poll(2) design, no lwp will ever need to grow bitmap of other lwp.
131  * This assumption is not true for /dev/poll; hence the need for extra
132  * locking.
133  *
134  * To allow more parallelism, each /dev/poll file descriptor (indexed by
135  * minor number) has its own lock. Since read (dpioctl) is a much more
136  * frequent operation than write, we want to allow multiple reads on same
137  * /dev/poll fd. However, we prevent writes from being starved by giving
138  * priority to write operation. Theoretically writes can starve reads as
139  * well. But in practical sense this is not important because (1) writes
140  * happens less often than reads, and (2) write operation defines the
141  * content of poll fd a cache set. If writes happens so often that they
142  * can starve reads, that means the cached set is very unstable. It may
143  * not make sense to read an unstable cache set anyway. Therefore, the
144  * writers starving readers case is not handled in this design.
145  */
146 
147 int
148 _init()
149 {
150 	int	error;
151 
152 	dptblsize = DEVPOLLSIZE;
153 	devpolltbl = kmem_zalloc(sizeof (caddr_t) * dptblsize, KM_SLEEP);
154 	mutex_init(&devpoll_lock, NULL, MUTEX_DEFAULT, NULL);
155 	devpoll_init = 1;
156 	if ((error = mod_install(&modlinkage)) != 0) {
157 		mutex_destroy(&devpoll_lock);
158 		kmem_free(devpolltbl, sizeof (caddr_t) * dptblsize);
159 		devpoll_init = 0;
160 	}
161 	return (error);
162 }
163 
164 int
165 _fini()
166 {
167 	int error;
168 
169 	if ((error = mod_remove(&modlinkage)) != 0) {
170 		return (error);
171 	}
172 	mutex_destroy(&devpoll_lock);
173 	kmem_free(devpolltbl, sizeof (caddr_t) * dptblsize);
174 	return (0);
175 }
176 
177 int
178 _info(struct modinfo *modinfop)
179 {
180 	return (mod_info(&modlinkage, modinfop));
181 }
182 
183 /*ARGSUSED*/
184 static int
185 dpattach(dev_info_t *devi, ddi_attach_cmd_t cmd)
186 {
187 	if (ddi_create_minor_node(devi, "poll", S_IFCHR, 0, DDI_PSEUDO, NULL)
188 	    == DDI_FAILURE) {
189 		ddi_remove_minor_node(devi, NULL);
190 		return (DDI_FAILURE);
191 	}
192 	dpdevi = devi;
193 	return (DDI_SUCCESS);
194 }
195 
196 static int
197 dpdetach(dev_info_t *devi, ddi_detach_cmd_t cmd)
198 {
199 	if (cmd != DDI_DETACH)
200 		return (DDI_FAILURE);
201 
202 	ddi_remove_minor_node(devi, NULL);
203 	return (DDI_SUCCESS);
204 }
205 
206 /* ARGSUSED */
207 static int
208 dpinfo(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
209 {
210 	int error;
211 
212 	switch (infocmd) {
213 	case DDI_INFO_DEVT2DEVINFO:
214 		*result = (void *)dpdevi;
215 		error = DDI_SUCCESS;
216 		break;
217 	case DDI_INFO_DEVT2INSTANCE:
218 		*result = (void *)0;
219 		error = DDI_SUCCESS;
220 		break;
221 	default:
222 		error = DDI_FAILURE;
223 	}
224 	return (error);
225 }
226 
227 /*
228  * dp_pcache_poll has similar logic to pcache_poll() in poll.c. The major
229  * differences are: (1) /dev/poll requires scanning the bitmap starting at
230  * where it was stopped last time, instead of always starting from 0,
231  * (2) since user may not have cleaned up the cached fds when they are
232  * closed, some polldats in cache may refer to closed or reused fds. We
233  * need to check for those cases.
234  *
235  * NOTE: Upon closing an fd, automatic poll cache cleanup is done for
236  *	 poll(2) caches but NOT for /dev/poll caches. So expect some
237  *	 stale entries!
238  */
239 static int
240 dp_pcache_poll(pollfd_t *pfdp, pollcache_t *pcp, nfds_t nfds, int *fdcntp)
241 {
242 	int		start, ostart, end;
243 	int		fdcnt, fd;
244 	boolean_t	done;
245 	file_t		*fp;
246 	short		revent;
247 	boolean_t	no_wrap;
248 	pollhead_t	*php;
249 	polldat_t	*pdp;
250 	int		error = 0;
251 
252 	ASSERT(MUTEX_HELD(&pcp->pc_lock));
253 	if (pcp->pc_bitmap == NULL) {
254 		/*
255 		 * No Need to search because no poll fd
256 		 * has been cached.
257 		 */
258 		return (error);
259 	}
260 retry:
261 	start = ostart = pcp->pc_mapstart;
262 	end = pcp->pc_mapend;
263 	php = NULL;
264 
265 	if (start == 0) {
266 		/*
267 		 * started from every begining, no need to wrap around.
268 		 */
269 		no_wrap = B_TRUE;
270 	} else {
271 		no_wrap = B_FALSE;
272 	}
273 	done = B_FALSE;
274 	fdcnt = 0;
275 	while ((fdcnt < nfds) && !done) {
276 		php = NULL;
277 		revent = 0;
278 		/*
279 		 * Examine the bit map in a circular fashion
280 		 * to avoid starvation. Always resume from
281 		 * last stop. Scan till end of the map. Then
282 		 * wrap around.
283 		 */
284 		fd = bt_getlowbit(pcp->pc_bitmap, start, end);
285 		ASSERT(fd <= end);
286 		if (fd >= 0) {
287 			if (fd == end) {
288 				if (no_wrap) {
289 					done = B_TRUE;
290 				} else {
291 					start = 0;
292 					end = ostart - 1;
293 					no_wrap = B_TRUE;
294 				}
295 			} else {
296 				start = fd + 1;
297 			}
298 			pdp = pcache_lookup_fd(pcp, fd);
299 repoll:
300 			ASSERT(pdp != NULL);
301 			ASSERT(pdp->pd_fd == fd);
302 			if (pdp->pd_fp == NULL) {
303 				/*
304 				 * The fd is POLLREMOVed. This fd is
305 				 * logically no longer cached. So move
306 				 * on to the next one.
307 				 */
308 				continue;
309 			}
310 			if ((fp = getf(fd)) == NULL) {
311 				/*
312 				 * The fd has been closed, but user has not
313 				 * done a POLLREMOVE on this fd yet. Instead
314 				 * of cleaning it here implicitly, we return
315 				 * POLLNVAL. This is consistent with poll(2)
316 				 * polling a closed fd. Hope this will remind
317 				 * user to do a POLLREMOVE.
318 				 */
319 				pfdp[fdcnt].fd = fd;
320 				pfdp[fdcnt].revents = POLLNVAL;
321 				fdcnt++;
322 				continue;
323 			}
324 			if (fp != pdp->pd_fp) {
325 				/*
326 				 * user is polling on a cached fd which was
327 				 * closed and then reused. Unfortunately
328 				 * there is no good way to inform user.
329 				 * If the file struct is also reused, we
330 				 * may not be able to detect the fd reuse
331 				 * at all.  As long as this does not
332 				 * cause system failure and/or memory leak,
333 				 * we will play along. Man page states if
334 				 * user does not clean up closed fds, polling
335 				 * results will be indeterministic.
336 				 *
337 				 * XXX - perhaps log the detection of fd
338 				 *	 reuse?
339 				 */
340 				pdp->pd_fp = fp;
341 			}
342 			/*
343 			 * XXX - pollrelock() logic needs to know which
344 			 * which pollcache lock to grab. It'd be a
345 			 * cleaner solution if we could pass pcp as
346 			 * an arguement in VOP_POLL interface instead
347 			 * of implicitly passing it using thread_t
348 			 * struct. On the other hand, changing VOP_POLL
349 			 * interface will require all driver/file system
350 			 * poll routine to change. May want to revisit
351 			 * the tradeoff later.
352 			 */
353 			curthread->t_pollcache = pcp;
354 			error = VOP_POLL(fp->f_vnode, pdp->pd_events, 0,
355 			    &revent, &php, NULL);
356 			curthread->t_pollcache = NULL;
357 			releasef(fd);
358 			if (error != 0) {
359 				break;
360 			}
361 			/*
362 			 * layered devices (e.g. console driver)
363 			 * may change the vnode and thus the pollhead
364 			 * pointer out from underneath us.
365 			 */
366 			if (php != NULL && pdp->pd_php != NULL &&
367 			    php != pdp->pd_php) {
368 				pollhead_delete(pdp->pd_php, pdp);
369 				pdp->pd_php = php;
370 				pollhead_insert(php, pdp);
371 				/*
372 				 * The bit should still be set.
373 				 */
374 				ASSERT(BT_TEST(pcp->pc_bitmap, fd));
375 				goto retry;
376 			}
377 
378 			if (revent != 0) {
379 				pfdp[fdcnt].fd = fd;
380 				pfdp[fdcnt].events = pdp->pd_events;
381 				pfdp[fdcnt].revents = revent;
382 				fdcnt++;
383 			} else if (php != NULL) {
384 				/*
385 				 * We clear a bit or cache a poll fd if
386 				 * the driver returns a poll head ptr,
387 				 * which is expected in the case of 0
388 				 * revents. Some buggy driver may return
389 				 * NULL php pointer with 0 revents. In
390 				 * this case, we just treat the driver as
391 				 * "noncachable" and not clearing the bit
392 				 * in bitmap.
393 				 */
394 				if ((pdp->pd_php != NULL) &&
395 				    ((pcp->pc_flag & T_POLLWAKE) == 0)) {
396 					BT_CLEAR(pcp->pc_bitmap, fd);
397 				}
398 				if (pdp->pd_php == NULL) {
399 					pollhead_insert(php, pdp);
400 					pdp->pd_php = php;
401 					/*
402 					 * An event of interest may have
403 					 * arrived between the VOP_POLL() and
404 					 * the pollhead_insert(); check again.
405 					 */
406 					goto repoll;
407 				}
408 			}
409 		} else {
410 			/*
411 			 * No bit set in the range. Check for wrap around.
412 			 */
413 			if (!no_wrap) {
414 				start = 0;
415 				end = ostart - 1;
416 				no_wrap = B_TRUE;
417 			} else {
418 				done = B_TRUE;
419 			}
420 		}
421 	}
422 
423 	if (!done) {
424 		pcp->pc_mapstart = start;
425 	}
426 	ASSERT(*fdcntp == 0);
427 	*fdcntp = fdcnt;
428 	return (error);
429 }
430 
431 /*ARGSUSED*/
432 static int
433 dpopen(dev_t *devp, int flag, int otyp, cred_t *credp)
434 {
435 	minor_t		minordev;
436 	dp_entry_t	*dpep;
437 	pollcache_t	*pcp;
438 
439 	ASSERT(devpoll_init);
440 	ASSERT(dptblsize <= MAXMIN);
441 	mutex_enter(&devpoll_lock);
442 	for (minordev = 0; minordev < dptblsize; minordev++) {
443 		if (devpolltbl[minordev] == NULL) {
444 			devpolltbl[minordev] = (dp_entry_t *)RESERVED;
445 			break;
446 		}
447 	}
448 	if (minordev == dptblsize) {
449 		dp_entry_t	**newtbl;
450 		size_t		oldsize;
451 
452 		/*
453 		 * Used up every entry in the existing devpoll table.
454 		 * Grow the table by DEVPOLLSIZE.
455 		 */
456 		if ((oldsize = dptblsize) >= MAXMIN) {
457 			mutex_exit(&devpoll_lock);
458 			return (ENXIO);
459 		}
460 		dptblsize += DEVPOLLSIZE;
461 		if (dptblsize > MAXMIN) {
462 			dptblsize = MAXMIN;
463 		}
464 		newtbl = kmem_zalloc(sizeof (caddr_t) * dptblsize, KM_SLEEP);
465 		bcopy(devpolltbl, newtbl, sizeof (caddr_t) * oldsize);
466 		kmem_free(devpolltbl, sizeof (caddr_t) * oldsize);
467 		devpolltbl = newtbl;
468 		devpolltbl[minordev] = (dp_entry_t *)RESERVED;
469 	}
470 	mutex_exit(&devpoll_lock);
471 
472 	dpep = kmem_zalloc(sizeof (dp_entry_t), KM_SLEEP);
473 	/*
474 	 * allocate a pollcache skeleton here. Delay allocating bitmap
475 	 * structures until dpwrite() time, since we don't know the
476 	 * optimal size yet.
477 	 */
478 	pcp = pcache_alloc();
479 	dpep->dpe_pcache = pcp;
480 	pcp->pc_pid = curproc->p_pid;
481 	*devp = makedevice(getmajor(*devp), minordev);  /* clone the driver */
482 	mutex_enter(&devpoll_lock);
483 	ASSERT(minordev < dptblsize);
484 	ASSERT(devpolltbl[minordev] == (dp_entry_t *)RESERVED);
485 	devpolltbl[minordev] = dpep;
486 	mutex_exit(&devpoll_lock);
487 	return (0);
488 }
489 
490 /*
491  * Write to dev/poll add/remove fd's to/from a cached poll fd set,
492  * or change poll events for a watched fd.
493  */
494 /*ARGSUSED*/
495 static int
496 dpwrite(dev_t dev, struct uio *uiop, cred_t *credp)
497 {
498 	minor_t		minor;
499 	dp_entry_t	*dpep;
500 	pollcache_t	*pcp;
501 	pollfd_t	*pollfdp, *pfdp;
502 	int		error;
503 	ssize_t		uiosize;
504 	nfds_t		pollfdnum;
505 	struct pollhead	*php = NULL;
506 	polldat_t	*pdp;
507 	int		fd;
508 	file_t		*fp;
509 
510 	minor = getminor(dev);
511 
512 	mutex_enter(&devpoll_lock);
513 	ASSERT(minor < dptblsize);
514 	dpep = devpolltbl[minor];
515 	ASSERT(dpep != NULL);
516 	mutex_exit(&devpoll_lock);
517 	pcp = dpep->dpe_pcache;
518 	if (curproc->p_pid != pcp->pc_pid) {
519 		return (EACCES);
520 	}
521 	uiosize = uiop->uio_resid;
522 	pollfdnum = uiosize / sizeof (pollfd_t);
523 	mutex_enter(&curproc->p_lock);
524 	if (pollfdnum > (uint_t)rctl_enforced_value(
525 	    rctlproc_legacy[RLIMIT_NOFILE], curproc->p_rctls, curproc)) {
526 		(void) rctl_action(rctlproc_legacy[RLIMIT_NOFILE],
527 		    curproc->p_rctls, curproc, RCA_SAFE);
528 		mutex_exit(&curproc->p_lock);
529 		return (set_errno(EINVAL));
530 	}
531 	mutex_exit(&curproc->p_lock);
532 	/*
533 	 * Copy in the pollfd array.  Walk through the array and add
534 	 * each polled fd to the cached set.
535 	 */
536 	pollfdp = kmem_alloc(uiosize, KM_SLEEP);
537 
538 	/*
539 	 * Although /dev/poll uses the write(2) interface to cache fds, it's
540 	 * not supposed to function as a seekable device. To prevent offset
541 	 * from growing and eventually exceed the maximum, reset the offset
542 	 * here for every call.
543 	 */
544 	uiop->uio_loffset = 0;
545 	if ((error = uiomove((caddr_t)pollfdp, uiosize, UIO_WRITE, uiop))
546 	    != 0) {
547 		kmem_free(pollfdp, uiosize);
548 		return (error);
549 	}
550 	/*
551 	 * We are about to enter the core portion of dpwrite(). Make sure this
552 	 * write has exclusive access in this portion of the code, i.e., no
553 	 * other writers in this code and no other readers in dpioctl.
554 	 */
555 	mutex_enter(&dpep->dpe_lock);
556 	dpep->dpe_writerwait++;
557 	while (dpep->dpe_refcnt != 0) {
558 		if (!cv_wait_sig_swap(&dpep->dpe_cv, &dpep->dpe_lock)) {
559 			dpep->dpe_writerwait--;
560 			mutex_exit(&dpep->dpe_lock);
561 			kmem_free(pollfdp, uiosize);
562 			return (set_errno(EINTR));
563 		}
564 	}
565 	dpep->dpe_writerwait--;
566 	dpep->dpe_flag |= DP_WRITER_PRESENT;
567 	dpep->dpe_refcnt++;
568 	mutex_exit(&dpep->dpe_lock);
569 
570 	mutex_enter(&pcp->pc_lock);
571 	if (pcp->pc_bitmap == NULL) {
572 		pcache_create(pcp, pollfdnum);
573 	}
574 	for (pfdp = pollfdp; pfdp < pollfdp + pollfdnum; pfdp++) {
575 		fd = pfdp->fd;
576 		if ((uint_t)fd >= P_FINFO(curproc)->fi_nfiles)
577 			continue;
578 		pdp = pcache_lookup_fd(pcp, fd);
579 		if (pfdp->events != POLLREMOVE) {
580 			if (pdp == NULL) {
581 				pdp = pcache_alloc_fd(0);
582 				pdp->pd_fd = fd;
583 				pdp->pd_pcache = pcp;
584 				pcache_insert_fd(pcp, pdp, pollfdnum);
585 			}
586 			ASSERT(pdp->pd_fd == fd);
587 			ASSERT(pdp->pd_pcache == pcp);
588 			if (fd >= pcp->pc_mapsize) {
589 				mutex_exit(&pcp->pc_lock);
590 				pcache_grow_map(pcp, fd);
591 				mutex_enter(&pcp->pc_lock);
592 			}
593 			if (fd > pcp->pc_mapend) {
594 				pcp->pc_mapend = fd;
595 			}
596 			if ((fp = getf(fd)) == NULL) {
597 				/*
598 				 * The fd is not valid. Since we can't pass
599 				 * this error back in the write() call, set
600 				 * the bit in bitmap to force DP_POLL ioctl
601 				 * to examine it.
602 				 */
603 				BT_SET(pcp->pc_bitmap, fd);
604 				pdp->pd_events |= pfdp->events;
605 				continue;
606 			}
607 			/*
608 			 * Don't do VOP_POLL for an already cached fd with
609 			 * same poll events.
610 			 */
611 			if ((pdp->pd_events == pfdp->events) &&
612 			    (pdp->pd_fp != NULL)) {
613 				/*
614 				 * the events are already cached
615 				 */
616 				releasef(fd);
617 				continue;
618 			}
619 
620 			/*
621 			 * do VOP_POLL and cache this poll fd.
622 			 */
623 			/*
624 			 * XXX - pollrelock() logic needs to know which
625 			 * which pollcache lock to grab. It'd be a
626 			 * cleaner solution if we could pass pcp as
627 			 * an arguement in VOP_POLL interface instead
628 			 * of implicitly passing it using thread_t
629 			 * struct. On the other hand, changing VOP_POLL
630 			 * interface will require all driver/file system
631 			 * poll routine to change. May want to revisit
632 			 * the tradeoff later.
633 			 */
634 			curthread->t_pollcache = pcp;
635 			error = VOP_POLL(fp->f_vnode, pfdp->events, 0,
636 			    &pfdp->revents, &php, NULL);
637 			curthread->t_pollcache = NULL;
638 			/*
639 			 * We always set the bit when this fd is cached;
640 			 * this forces the first DP_POLL to poll this fd.
641 			 * Real performance gain comes from subsequent
642 			 * DP_POLL.  We also attempt a pollhead_insert();
643 			 * if it's not possible, we'll do it in dpioctl().
644 			 */
645 			BT_SET(pcp->pc_bitmap, fd);
646 			if (error != 0) {
647 				releasef(fd);
648 				break;
649 			}
650 			pdp->pd_fp = fp;
651 			pdp->pd_events |= pfdp->events;
652 			if (php != NULL) {
653 				if (pdp->pd_php == NULL) {
654 					pollhead_insert(php, pdp);
655 					pdp->pd_php = php;
656 				} else {
657 					if (pdp->pd_php != php) {
658 						pollhead_delete(pdp->pd_php,
659 						    pdp);
660 						pollhead_insert(php, pdp);
661 						pdp->pd_php = php;
662 					}
663 				}
664 
665 			}
666 			releasef(fd);
667 		} else {
668 			if (pdp == NULL) {
669 				continue;
670 			}
671 			ASSERT(pdp->pd_fd == fd);
672 			pdp->pd_fp = NULL;
673 			pdp->pd_events = 0;
674 			ASSERT(pdp->pd_thread == NULL);
675 			if (pdp->pd_php != NULL) {
676 				pollhead_delete(pdp->pd_php, pdp);
677 				pdp->pd_php = NULL;
678 			}
679 			BT_CLEAR(pcp->pc_bitmap, fd);
680 		}
681 	}
682 	mutex_exit(&pcp->pc_lock);
683 	mutex_enter(&dpep->dpe_lock);
684 	dpep->dpe_flag &= ~DP_WRITER_PRESENT;
685 	ASSERT(dpep->dpe_refcnt == 1);
686 	dpep->dpe_refcnt--;
687 	cv_broadcast(&dpep->dpe_cv);
688 	mutex_exit(&dpep->dpe_lock);
689 	kmem_free(pollfdp, uiosize);
690 	return (error);
691 }
692 
693 /*ARGSUSED*/
694 static int
695 dpioctl(dev_t dev, int cmd, intptr_t arg, int mode, cred_t *credp, int *rvalp)
696 {
697 	minor_t		minor;
698 	dp_entry_t	*dpep;
699 	pollcache_t	*pcp;
700 	hrtime_t	now;
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 */
706 		now = gethrtime();
707 	}
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 		hrtime_t	deadline = 0;
737 
738 		STRUCT_INIT(dvpoll, mode);
739 		error = copyin((caddr_t)arg, STRUCT_BUF(dvpoll),
740 		    STRUCT_SIZE(dvpoll));
741 		if (error) {
742 			DP_REFRELE(dpep);
743 			return (EFAULT);
744 		}
745 
746 		deadline = STRUCT_FGET(dvpoll, dp_timeout);
747 		if (deadline > 0) {
748 			/*
749 			 * Convert the deadline from relative milliseconds
750 			 * to absolute nanoseconds.  They must wait for at
751 			 * least a tick.
752 			 */
753 			deadline = deadline * NANOSEC / MILLISEC;
754 			deadline = MAX(deadline, nsec_per_tick);
755 			deadline += 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 (deadline == 0)
766 				return (0);
767 			mutex_enter(&curthread->t_delay_lock);
768 			while ((error =
769 			    cv_timedwait_sig_hrtime(&curthread->t_delay_cv,
770 			    &curthread->t_delay_lock, deadline)) > 0)
771 				continue;
772 			mutex_exit(&curthread->t_delay_lock);
773 			return (error == 0 ? EINTR : 0);
774 		}
775 
776 		/*
777 		 * XXX It would be nice not to have to alloc each time, but it
778 		 * requires another per thread structure hook. This can be
779 		 * implemented later if data suggests that it's necessary.
780 		 */
781 		if ((ps = curthread->t_pollstate) == NULL) {
782 			curthread->t_pollstate = pollstate_create();
783 			ps = curthread->t_pollstate;
784 		}
785 		if (ps->ps_dpbufsize < nfds) {
786 			struct proc *p = ttoproc(curthread);
787 			/*
788 			 * The maximum size should be no large than
789 			 * current maximum open file count.
790 			 */
791 			mutex_enter(&p->p_lock);
792 			if (nfds > p->p_fno_ctl) {
793 				mutex_exit(&p->p_lock);
794 				DP_REFRELE(dpep);
795 				return (EINVAL);
796 			}
797 			mutex_exit(&p->p_lock);
798 			kmem_free(ps->ps_dpbuf, sizeof (pollfd_t) *
799 			    ps->ps_dpbufsize);
800 			ps->ps_dpbuf = kmem_zalloc(sizeof (pollfd_t) *
801 			    nfds, KM_SLEEP);
802 			ps->ps_dpbufsize = nfds;
803 		}
804 
805 		mutex_enter(&pcp->pc_lock);
806 		for (;;) {
807 			pcp->pc_flag = 0;
808 			error = dp_pcache_poll(ps->ps_dpbuf, pcp, nfds, &fdcnt);
809 			if (fdcnt > 0 || error != 0)
810 				break;
811 
812 			/*
813 			 * A pollwake has happened since we polled cache.
814 			 */
815 			if (pcp->pc_flag & T_POLLWAKE)
816 				continue;
817 
818 			/*
819 			 * Sleep until we are notified, signaled, or timed out.
820 			 */
821 			if (deadline == 0) {
822 				/* immediate timeout; do not check signals */
823 				break;
824 			}
825 			error = cv_timedwait_sig_hrtime(&pcp->pc_cv,
826 			    &pcp->pc_lock, deadline);
827 			/*
828 			 * If we were awakened by a signal or timeout
829 			 * then break the loop, else poll again.
830 			 */
831 			if (error <= 0) {
832 				error = (error == 0) ? EINTR : 0;
833 				break;
834 			} else {
835 				error = 0;
836 			}
837 		}
838 		mutex_exit(&pcp->pc_lock);
839 
840 		if (error == 0 && fdcnt > 0) {
841 			if (copyout(ps->ps_dpbuf, STRUCT_FGETP(dvpoll,
842 			    dp_fds), sizeof (pollfd_t) * fdcnt)) {
843 				DP_REFRELE(dpep);
844 				return (EFAULT);
845 			}
846 			*rvalp = fdcnt;
847 		}
848 		break;
849 	}
850 
851 	case	DP_ISPOLLED:
852 	{
853 		pollfd_t	pollfd;
854 		polldat_t	*pdp;
855 
856 		STRUCT_INIT(dvpoll, mode);
857 		error = copyin((caddr_t)arg, &pollfd, sizeof (pollfd_t));
858 		if (error) {
859 			DP_REFRELE(dpep);
860 			return (EFAULT);
861 		}
862 		mutex_enter(&pcp->pc_lock);
863 		if (pcp->pc_hash == NULL) {
864 			/*
865 			 * No Need to search because no poll fd
866 			 * has been cached.
867 			 */
868 			mutex_exit(&pcp->pc_lock);
869 			DP_REFRELE(dpep);
870 			return (0);
871 		}
872 		if (pollfd.fd < 0) {
873 			mutex_exit(&pcp->pc_lock);
874 			break;
875 		}
876 		pdp = pcache_lookup_fd(pcp, pollfd.fd);
877 		if ((pdp != NULL) && (pdp->pd_fd == pollfd.fd) &&
878 		    (pdp->pd_fp != NULL)) {
879 			pollfd.revents = pdp->pd_events;
880 			if (copyout(&pollfd, (caddr_t)arg, sizeof (pollfd_t))) {
881 				mutex_exit(&pcp->pc_lock);
882 				DP_REFRELE(dpep);
883 				return (EFAULT);
884 			}
885 			*rvalp = 1;
886 		}
887 		mutex_exit(&pcp->pc_lock);
888 		break;
889 	}
890 
891 	default:
892 		DP_REFRELE(dpep);
893 		return (EINVAL);
894 	}
895 	DP_REFRELE(dpep);
896 	return (error);
897 }
898 
899 /*ARGSUSED*/
900 static int
901 dppoll(dev_t dev, short events, int anyyet, short *reventsp,
902     struct pollhead **phpp)
903 {
904 	/*
905 	 * Polling on a /dev/poll fd is not fully supported yet.
906 	 */
907 	*reventsp = POLLERR;
908 	return (0);
909 }
910 
911 /*
912  * devpoll close should do enough clean up before the pollcache is deleted,
913  * i.e., it should ensure no one still references the pollcache later.
914  * There is no "permission" check in here. Any process having the last
915  * reference of this /dev/poll fd can close.
916  */
917 /*ARGSUSED*/
918 static int
919 dpclose(dev_t dev, int flag, int otyp, cred_t *credp)
920 {
921 	minor_t		minor;
922 	dp_entry_t	*dpep;
923 	pollcache_t	*pcp;
924 	int		i;
925 	polldat_t	**hashtbl;
926 	polldat_t	*pdp;
927 
928 	minor = getminor(dev);
929 
930 	mutex_enter(&devpoll_lock);
931 	dpep = devpolltbl[minor];
932 	ASSERT(dpep != NULL);
933 	devpolltbl[minor] = NULL;
934 	mutex_exit(&devpoll_lock);
935 	pcp = dpep->dpe_pcache;
936 	ASSERT(pcp != NULL);
937 	/*
938 	 * At this point, no other lwp can access this pollcache via the
939 	 * /dev/poll fd. This pollcache is going away, so do the clean
940 	 * up without the pc_lock.
941 	 */
942 	hashtbl = pcp->pc_hash;
943 	for (i = 0; i < pcp->pc_hashsize; i++) {
944 		for (pdp = hashtbl[i]; pdp; pdp = pdp->pd_hashnext) {
945 			if (pdp->pd_php != NULL) {
946 				pollhead_delete(pdp->pd_php, pdp);
947 				pdp->pd_php = NULL;
948 				pdp->pd_fp = NULL;
949 			}
950 		}
951 	}
952 	/*
953 	 * pollwakeup() may still interact with this pollcache. Wait until
954 	 * it is done.
955 	 */
956 	mutex_enter(&pcp->pc_no_exit);
957 	ASSERT(pcp->pc_busy >= 0);
958 	while (pcp->pc_busy > 0)
959 		cv_wait(&pcp->pc_busy_cv, &pcp->pc_no_exit);
960 	mutex_exit(&pcp->pc_no_exit);
961 	pcache_destroy(pcp);
962 	ASSERT(dpep->dpe_refcnt == 0);
963 	kmem_free(dpep, sizeof (dp_entry_t));
964 	return (0);
965 }
966