Print this page
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/io/devpoll.c
+++ new/usr/src/uts/common/io/devpoll.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
13 13 * When distributing Covered Code, include this CDDL HEADER in each
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
19 19 * CDDL HEADER END
20 20 */
21 21 /*
22 22 * Copyright 2008 Sun Microsystems, Inc. All rights reserved.
23 23 * Use is subject to license terms.
24 24 */
25 25
26 26 /*
27 27 * Copyright (c) 2012 by Delphix. All rights reserved.
28 28 * Copyright 2016 Joyent, Inc.
29 29 */
30 30
31 31 #include <sys/types.h>
32 32 #include <sys/devops.h>
33 33 #include <sys/conf.h>
34 34 #include <sys/modctl.h>
35 35 #include <sys/sunddi.h>
36 36 #include <sys/stat.h>
37 37 #include <sys/poll_impl.h>
38 38 #include <sys/errno.h>
39 39 #include <sys/kmem.h>
40 40 #include <sys/mkdev.h>
41 41 #include <sys/debug.h>
42 42 #include <sys/file.h>
43 43 #include <sys/sysmacros.h>
44 44 #include <sys/systm.h>
45 45 #include <sys/bitmap.h>
46 46 #include <sys/devpoll.h>
47 47 #include <sys/rctl.h>
48 48 #include <sys/resource.h>
49 49 #include <sys/schedctl.h>
50 50 #include <sys/epoll.h>
51 51
52 52 #define RESERVED 1
53 53
54 54 /* local data struct */
55 55 static dp_entry_t **devpolltbl; /* dev poll entries */
56 56 static size_t dptblsize;
57 57
58 58 static kmutex_t devpoll_lock; /* lock protecting dev tbl */
59 59 int devpoll_init; /* is /dev/poll initialized already */
60 60
61 61 /* device local functions */
62 62
63 63 static int dpopen(dev_t *devp, int flag, int otyp, cred_t *credp);
64 64 static int dpwrite(dev_t dev, struct uio *uiop, cred_t *credp);
65 65 static int dpioctl(dev_t dev, int cmd, intptr_t arg, int mode, cred_t *credp,
66 66 int *rvalp);
67 67 static int dppoll(dev_t dev, short events, int anyyet, short *reventsp,
68 68 struct pollhead **phpp);
69 69 static int dpclose(dev_t dev, int flag, int otyp, cred_t *credp);
70 70 static dev_info_t *dpdevi;
71 71
72 72
73 73 static struct cb_ops dp_cb_ops = {
74 74 dpopen, /* open */
75 75 dpclose, /* close */
76 76 nodev, /* strategy */
77 77 nodev, /* print */
78 78 nodev, /* dump */
79 79 nodev, /* read */
80 80 dpwrite, /* write */
81 81 dpioctl, /* ioctl */
82 82 nodev, /* devmap */
83 83 nodev, /* mmap */
84 84 nodev, /* segmap */
85 85 dppoll, /* poll */
86 86 ddi_prop_op, /* prop_op */
87 87 (struct streamtab *)0, /* streamtab */
88 88 D_MP, /* flags */
89 89 CB_REV, /* cb_ops revision */
90 90 nodev, /* aread */
91 91 nodev /* awrite */
92 92 };
93 93
94 94 static int dpattach(dev_info_t *, ddi_attach_cmd_t);
95 95 static int dpdetach(dev_info_t *, ddi_detach_cmd_t);
96 96 static int dpinfo(dev_info_t *, ddi_info_cmd_t, void *, void **);
97 97
98 98 static struct dev_ops dp_ops = {
99 99 DEVO_REV, /* devo_rev */
100 100 0, /* refcnt */
101 101 dpinfo, /* info */
102 102 nulldev, /* identify */
103 103 nulldev, /* probe */
104 104 dpattach, /* attach */
105 105 dpdetach, /* detach */
106 106 nodev, /* reset */
107 107 &dp_cb_ops, /* driver operations */
108 108 (struct bus_ops *)NULL, /* bus operations */
109 109 nulldev, /* power */
110 110 ddi_quiesce_not_needed, /* quiesce */
111 111 };
112 112
113 113
114 114 static struct modldrv modldrv = {
115 115 &mod_driverops, /* type of module - a driver */
116 116 "/dev/poll driver",
117 117 &dp_ops,
118 118 };
119 119
120 120 static struct modlinkage modlinkage = {
121 121 MODREV_1,
122 122 (void *)&modldrv,
123 123 NULL
124 124 };
125 125
126 126 static void pcachelink_assoc(pollcache_t *, pollcache_t *);
127 127 static void pcachelink_mark_stale(pollcache_t *);
128 128 static void pcachelink_purge_stale(pollcache_t *);
129 129 static void pcachelink_purge_all(pollcache_t *);
130 130
131 131
132 132 /*
133 133 * Locking Design
134 134 *
135 135 * The /dev/poll driver shares most of its code with poll sys call whose
136 136 * code is in common/syscall/poll.c. In poll(2) design, the pollcache
137 137 * structure is per lwp. An implicit assumption is made there that some
138 138 * portion of pollcache will never be touched by other lwps. E.g., in
139 139 * poll(2) design, no lwp will ever need to grow bitmap of other lwp.
140 140 * This assumption is not true for /dev/poll; hence the need for extra
141 141 * locking.
142 142 *
143 143 * To allow more parallelism, each /dev/poll file descriptor (indexed by
144 144 * minor number) has its own lock. Since read (dpioctl) is a much more
145 145 * frequent operation than write, we want to allow multiple reads on same
146 146 * /dev/poll fd. However, we prevent writes from being starved by giving
147 147 * priority to write operation. Theoretically writes can starve reads as
148 148 * well. But in practical sense this is not important because (1) writes
149 149 * happens less often than reads, and (2) write operation defines the
150 150 * content of poll fd a cache set. If writes happens so often that they
151 151 * can starve reads, that means the cached set is very unstable. It may
152 152 * not make sense to read an unstable cache set anyway. Therefore, the
153 153 * writers starving readers case is not handled in this design.
154 154 */
155 155
156 156 int
157 157 _init()
158 158 {
159 159 int error;
160 160
161 161 dptblsize = DEVPOLLSIZE;
162 162 devpolltbl = kmem_zalloc(sizeof (caddr_t) * dptblsize, KM_SLEEP);
163 163 mutex_init(&devpoll_lock, NULL, MUTEX_DEFAULT, NULL);
164 164 devpoll_init = 1;
165 165 if ((error = mod_install(&modlinkage)) != 0) {
166 166 kmem_free(devpolltbl, sizeof (caddr_t) * dptblsize);
167 167 devpoll_init = 0;
168 168 }
169 169 return (error);
170 170 }
171 171
172 172 int
173 173 _fini()
174 174 {
175 175 int error;
176 176
177 177 if ((error = mod_remove(&modlinkage)) != 0) {
178 178 return (error);
179 179 }
180 180 mutex_destroy(&devpoll_lock);
181 181 kmem_free(devpolltbl, sizeof (caddr_t) * dptblsize);
182 182 return (0);
183 183 }
184 184
185 185 int
186 186 _info(struct modinfo *modinfop)
187 187 {
188 188 return (mod_info(&modlinkage, modinfop));
189 189 }
190 190
191 191 /*ARGSUSED*/
192 192 static int
193 193 dpattach(dev_info_t *devi, ddi_attach_cmd_t cmd)
194 194 {
195 195 if (ddi_create_minor_node(devi, "poll", S_IFCHR, 0, DDI_PSEUDO, NULL)
196 196 == DDI_FAILURE) {
197 197 ddi_remove_minor_node(devi, NULL);
198 198 return (DDI_FAILURE);
199 199 }
200 200 dpdevi = devi;
201 201 return (DDI_SUCCESS);
202 202 }
203 203
204 204 static int
205 205 dpdetach(dev_info_t *devi, ddi_detach_cmd_t cmd)
206 206 {
207 207 if (cmd != DDI_DETACH)
208 208 return (DDI_FAILURE);
209 209
210 210 ddi_remove_minor_node(devi, NULL);
211 211 return (DDI_SUCCESS);
212 212 }
213 213
214 214 /* ARGSUSED */
215 215 static int
216 216 dpinfo(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
217 217 {
218 218 int error;
219 219
220 220 switch (infocmd) {
221 221 case DDI_INFO_DEVT2DEVINFO:
222 222 *result = (void *)dpdevi;
223 223 error = DDI_SUCCESS;
224 224 break;
225 225 case DDI_INFO_DEVT2INSTANCE:
226 226 *result = (void *)0;
227 227 error = DDI_SUCCESS;
228 228 break;
229 229 default:
230 230 error = DDI_FAILURE;
231 231 }
232 232 return (error);
233 233 }
234 234
235 235 /*
236 236 * dp_pcache_poll has similar logic to pcache_poll() in poll.c. The major
237 237 * differences are: (1) /dev/poll requires scanning the bitmap starting at
238 238 * where it was stopped last time, instead of always starting from 0,
239 239 * (2) since user may not have cleaned up the cached fds when they are
240 240 * closed, some polldats in cache may refer to closed or reused fds. We
241 241 * need to check for those cases.
242 242 *
243 243 * NOTE: Upon closing an fd, automatic poll cache cleanup is done for
244 244 * poll(2) caches but NOT for /dev/poll caches. So expect some
245 245 * stale entries!
246 246 */
247 247 static int
248 248 dp_pcache_poll(dp_entry_t *dpep, void *dpbuf,
249 249 pollcache_t *pcp, nfds_t nfds, int *fdcntp)
250 250 {
251 251 int start, ostart, end;
252 252 int fdcnt, fd;
253 253 boolean_t done;
254 254 file_t *fp;
255 255 short revent;
256 256 boolean_t no_wrap;
257 257 pollhead_t *php;
258 258 polldat_t *pdp;
259 259 pollfd_t *pfdp;
260 260 epoll_event_t *epoll;
261 261 int error = 0;
262 262 short mask = POLLRDHUP | POLLWRBAND;
263 263 boolean_t is_epoll = (dpep->dpe_flag & DP_ISEPOLLCOMPAT) != 0;
264 264
265 265 ASSERT(MUTEX_HELD(&pcp->pc_lock));
266 266 if (pcp->pc_bitmap == NULL) {
267 267 /*
268 268 * No Need to search because no poll fd
269 269 * has been cached.
270 270 */
271 271 return (error);
272 272 }
273 273
274 274 if (is_epoll) {
275 275 pfdp = NULL;
276 276 epoll = (epoll_event_t *)dpbuf;
277 277 } else {
278 278 pfdp = (pollfd_t *)dpbuf;
279 279 epoll = NULL;
280 280 }
281 281 retry:
282 282 start = ostart = pcp->pc_mapstart;
283 283 end = pcp->pc_mapend;
284 284 php = NULL;
285 285
286 286 if (start == 0) {
287 287 /*
288 288 * started from every begining, no need to wrap around.
289 289 */
290 290 no_wrap = B_TRUE;
291 291 } else {
292 292 no_wrap = B_FALSE;
293 293 }
294 294 done = B_FALSE;
295 295 fdcnt = 0;
296 296 while ((fdcnt < nfds) && !done) {
297 297 php = NULL;
298 298 revent = 0;
299 299 /*
300 300 * Examine the bit map in a circular fashion
301 301 * to avoid starvation. Always resume from
302 302 * last stop. Scan till end of the map. Then
303 303 * wrap around.
304 304 */
305 305 fd = bt_getlowbit(pcp->pc_bitmap, start, end);
306 306 ASSERT(fd <= end);
307 307 if (fd >= 0) {
308 308 if (fd == end) {
309 309 if (no_wrap) {
310 310 done = B_TRUE;
311 311 } else {
312 312 start = 0;
313 313 end = ostart - 1;
314 314 no_wrap = B_TRUE;
315 315 }
316 316 } else {
317 317 start = fd + 1;
318 318 }
319 319 pdp = pcache_lookup_fd(pcp, fd);
320 320 repoll:
321 321 ASSERT(pdp != NULL);
322 322 ASSERT(pdp->pd_fd == fd);
323 323 if (pdp->pd_fp == NULL) {
324 324 /*
325 325 * The fd is POLLREMOVed. This fd is
326 326 * logically no longer cached. So move
327 327 * on to the next one.
328 328 */
329 329 continue;
330 330 }
331 331 if ((fp = getf(fd)) == NULL) {
332 332 /*
333 333 * The fd has been closed, but user has not
334 334 * done a POLLREMOVE on this fd yet. Instead
335 335 * of cleaning it here implicitly, we return
336 336 * POLLNVAL. This is consistent with poll(2)
337 337 * polling a closed fd. Hope this will remind
338 338 * user to do a POLLREMOVE.
339 339 */
340 340 if (!is_epoll && pfdp != NULL) {
341 341 pfdp[fdcnt].fd = fd;
342 342 pfdp[fdcnt].revents = POLLNVAL;
343 343 fdcnt++;
344 344 continue;
345 345 }
346 346
347 347 /*
348 348 * In the epoll compatibility case, we actually
349 349 * perform the implicit removal to remain
350 350 * closer to the epoll semantics.
351 351 */
352 352 if (is_epoll) {
353 353 pdp->pd_fp = NULL;
354 354 pdp->pd_events = 0;
355 355
356 356 if (pdp->pd_php != NULL) {
357 357 pollhead_delete(pdp->pd_php,
358 358 pdp);
359 359 pdp->pd_php = NULL;
360 360 }
361 361
362 362 BT_CLEAR(pcp->pc_bitmap, fd);
363 363 continue;
364 364 }
365 365 }
366 366
367 367 if (fp != pdp->pd_fp) {
368 368 /*
369 369 * user is polling on a cached fd which was
370 370 * closed and then reused. Unfortunately
371 371 * there is no good way to inform user.
372 372 * If the file struct is also reused, we
373 373 * may not be able to detect the fd reuse
374 374 * at all. As long as this does not
375 375 * cause system failure and/or memory leak,
376 376 * we will play along. Man page states if
377 377 * user does not clean up closed fds, polling
378 378 * results will be indeterministic.
379 379 *
380 380 * XXX - perhaps log the detection of fd
381 381 * reuse?
382 382 */
383 383 pdp->pd_fp = fp;
384 384 }
385 385 /*
386 386 * XXX - pollrelock() logic needs to know which
387 387 * which pollcache lock to grab. It'd be a
388 388 * cleaner solution if we could pass pcp as
389 389 * an arguement in VOP_POLL interface instead
390 390 * of implicitly passing it using thread_t
391 391 * struct. On the other hand, changing VOP_POLL
392 392 * interface will require all driver/file system
393 393 * poll routine to change. May want to revisit
394 394 * the tradeoff later.
395 395 */
396 396 curthread->t_pollcache = pcp;
397 397 error = VOP_POLL(fp->f_vnode, pdp->pd_events, 0,
398 398 &revent, &php, NULL);
399 399 curthread->t_pollcache = NULL;
400 400 releasef(fd);
401 401 if (error != 0) {
402 402 break;
403 403 }
404 404
405 405 /*
406 406 * layered devices (e.g. console driver)
407 407 * may change the vnode and thus the pollhead
408 408 * pointer out from underneath us.
409 409 */
410 410 if (php != NULL && pdp->pd_php != NULL &&
411 411 php != pdp->pd_php) {
412 412 pollhead_delete(pdp->pd_php, pdp);
413 413 pdp->pd_php = php;
414 414 pollhead_insert(php, pdp);
415 415 /*
416 416 * The bit should still be set.
417 417 */
418 418 ASSERT(BT_TEST(pcp->pc_bitmap, fd));
419 419 goto retry;
420 420 }
421 421
422 422 if (revent != 0) {
423 423 if (pfdp != NULL) {
424 424 pfdp[fdcnt].fd = fd;
425 425 pfdp[fdcnt].events = pdp->pd_events;
426 426 pfdp[fdcnt].revents = revent;
427 427 } else if (epoll != NULL) {
428 428 epoll_event_t *ep = &epoll[fdcnt];
429 429
430 430 ASSERT(epoll != NULL);
431 431 ep->data.u64 = pdp->pd_epolldata;
432 432
433 433 /*
434 434 * If any of the event bits are set for
435 435 * which poll and epoll representations
436 436 * differ, swizzle in the native epoll
437 437 * values.
438 438 */
439 439 if (revent & mask) {
440 440 ep->events = (revent & ~mask) |
441 441 ((revent & POLLRDHUP) ?
442 442 EPOLLRDHUP : 0) |
443 443 ((revent & POLLWRBAND) ?
444 444 EPOLLWRBAND : 0);
445 445 } else {
446 446 ep->events = revent;
447 447 }
448 448
449 449 /*
450 450 * We define POLLWRNORM to be POLLOUT,
451 451 * but epoll has separate definitions
452 452 * for them; if POLLOUT is set and the
453 453 * user has asked for EPOLLWRNORM, set
454 454 * that as well.
455 455 */
456 456 if ((revent & POLLOUT) &&
457 457 (pdp->pd_events & EPOLLWRNORM)) {
458 458 ep->events |= EPOLLWRNORM;
459 459 }
460 460 } else {
461 461 pollstate_t *ps =
462 462 curthread->t_pollstate;
463 463 /*
464 464 * The devpoll handle itself is being
465 465 * polled. Notify the caller of any
466 466 * readable event(s), leaving as much
467 467 * state as possible untouched.
468 468 */
469 469 VERIFY(fdcnt == 0);
470 470 VERIFY(ps != NULL);
471 471
472 472 /*
473 473 * If a call to pollunlock() fails
474 474 * during VOP_POLL, skip over the fd
475 475 * and continue polling.
476 476 *
477 477 * Otherwise, report that there is an
478 478 * event pending.
479 479 */
480 480 if ((ps->ps_flags & POLLSTATE_ULFAIL)
481 481 != 0) {
482 482 ps->ps_flags &=
483 483 ~POLLSTATE_ULFAIL;
484 484 continue;
485 485 } else {
486 486 fdcnt++;
487 487 break;
488 488 }
489 489 }
490 490
491 491 /*
492 492 * If POLLET is set, clear the bit in the
493 493 * bitmap -- which effectively latches the
494 494 * edge on a pollwakeup() from the driver.
495 495 */
496 496 if (pdp->pd_events & POLLET)
497 497 BT_CLEAR(pcp->pc_bitmap, fd);
498 498
499 499 /*
500 500 * If POLLONESHOT is set, perform the implicit
501 501 * POLLREMOVE.
502 502 */
503 503 if (pdp->pd_events & POLLONESHOT) {
504 504 pdp->pd_fp = NULL;
505 505 pdp->pd_events = 0;
506 506
507 507 if (pdp->pd_php != NULL) {
508 508 pollhead_delete(pdp->pd_php,
509 509 pdp);
510 510 pdp->pd_php = NULL;
511 511 }
512 512
513 513 BT_CLEAR(pcp->pc_bitmap, fd);
514 514 }
515 515
516 516 fdcnt++;
517 517 } else if (php != NULL) {
518 518 /*
519 519 * We clear a bit or cache a poll fd if
520 520 * the driver returns a poll head ptr,
521 521 * which is expected in the case of 0
522 522 * revents. Some buggy driver may return
523 523 * NULL php pointer with 0 revents. In
524 524 * this case, we just treat the driver as
525 525 * "noncachable" and not clearing the bit
526 526 * in bitmap.
527 527 */
528 528 if ((pdp->pd_php != NULL) &&
529 529 ((pcp->pc_flag & PC_POLLWAKE) == 0)) {
530 530 BT_CLEAR(pcp->pc_bitmap, fd);
531 531 }
532 532 if (pdp->pd_php == NULL) {
533 533 pollhead_insert(php, pdp);
534 534 pdp->pd_php = php;
535 535 /*
536 536 * An event of interest may have
537 537 * arrived between the VOP_POLL() and
538 538 * the pollhead_insert(); check again.
539 539 */
540 540 goto repoll;
541 541 }
542 542 }
543 543 } else {
544 544 /*
545 545 * No bit set in the range. Check for wrap around.
546 546 */
547 547 if (!no_wrap) {
548 548 start = 0;
549 549 end = ostart - 1;
550 550 no_wrap = B_TRUE;
551 551 } else {
552 552 done = B_TRUE;
553 553 }
554 554 }
555 555 }
556 556
557 557 if (!done) {
558 558 pcp->pc_mapstart = start;
559 559 }
560 560 ASSERT(*fdcntp == 0);
561 561 *fdcntp = fdcnt;
562 562 return (error);
563 563 }
564 564
565 565 /*ARGSUSED*/
566 566 static int
567 567 dpopen(dev_t *devp, int flag, int otyp, cred_t *credp)
568 568 {
569 569 minor_t minordev;
570 570 dp_entry_t *dpep;
571 571 pollcache_t *pcp;
572 572
573 573 ASSERT(devpoll_init);
574 574 ASSERT(dptblsize <= MAXMIN);
575 575 mutex_enter(&devpoll_lock);
576 576 for (minordev = 0; minordev < dptblsize; minordev++) {
577 577 if (devpolltbl[minordev] == NULL) {
578 578 devpolltbl[minordev] = (dp_entry_t *)RESERVED;
579 579 break;
580 580 }
581 581 }
582 582 if (minordev == dptblsize) {
583 583 dp_entry_t **newtbl;
584 584 size_t oldsize;
585 585
586 586 /*
587 587 * Used up every entry in the existing devpoll table.
588 588 * Grow the table by DEVPOLLSIZE.
589 589 */
590 590 if ((oldsize = dptblsize) >= MAXMIN) {
591 591 mutex_exit(&devpoll_lock);
592 592 return (ENXIO);
593 593 }
594 594 dptblsize += DEVPOLLSIZE;
595 595 if (dptblsize > MAXMIN) {
596 596 dptblsize = MAXMIN;
597 597 }
598 598 newtbl = kmem_zalloc(sizeof (caddr_t) * dptblsize, KM_SLEEP);
599 599 bcopy(devpolltbl, newtbl, sizeof (caddr_t) * oldsize);
600 600 kmem_free(devpolltbl, sizeof (caddr_t) * oldsize);
601 601 devpolltbl = newtbl;
602 602 devpolltbl[minordev] = (dp_entry_t *)RESERVED;
603 603 }
604 604 mutex_exit(&devpoll_lock);
605 605
606 606 dpep = kmem_zalloc(sizeof (dp_entry_t), KM_SLEEP);
607 607 /*
608 608 * allocate a pollcache skeleton here. Delay allocating bitmap
609 609 * structures until dpwrite() time, since we don't know the
610 610 * optimal size yet. We also delay setting the pid until either
611 611 * dpwrite() or attempt to poll on the instance, allowing parents
612 612 * to create instances of /dev/poll for their children. (In the
613 613 * epoll compatibility case, this check isn't performed to maintain
614 614 * semantic compatibility.)
615 615 */
616 616 pcp = pcache_alloc();
617 617 dpep->dpe_pcache = pcp;
618 618 pcp->pc_pid = -1;
619 619 *devp = makedevice(getmajor(*devp), minordev); /* clone the driver */
620 620 mutex_enter(&devpoll_lock);
621 621 ASSERT(minordev < dptblsize);
622 622 ASSERT(devpolltbl[minordev] == (dp_entry_t *)RESERVED);
623 623 devpolltbl[minordev] = dpep;
624 624 mutex_exit(&devpoll_lock);
625 625 return (0);
626 626 }
627 627
628 628 /*
629 629 * Write to dev/poll add/remove fd's to/from a cached poll fd set,
630 630 * or change poll events for a watched fd.
631 631 */
632 632 /*ARGSUSED*/
633 633 static int
634 634 dpwrite(dev_t dev, struct uio *uiop, cred_t *credp)
635 635 {
636 636 minor_t minor;
637 637 dp_entry_t *dpep;
638 638 pollcache_t *pcp;
639 639 pollfd_t *pollfdp, *pfdp;
640 640 dvpoll_epollfd_t *epfdp;
641 641 uintptr_t limit;
642 642 int error, size;
643 643 ssize_t uiosize;
644 644 size_t copysize;
645 645 nfds_t pollfdnum;
646 646 struct pollhead *php = NULL;
647 647 polldat_t *pdp;
648 648 int fd;
649 649 file_t *fp;
650 650 boolean_t is_epoll, fds_added = B_FALSE;
651 651
652 652 minor = getminor(dev);
653 653
654 654 mutex_enter(&devpoll_lock);
655 655 ASSERT(minor < dptblsize);
656 656 dpep = devpolltbl[minor];
657 657 ASSERT(dpep != NULL);
658 658 mutex_exit(&devpoll_lock);
659 659
660 660 mutex_enter(&dpep->dpe_lock);
661 661 pcp = dpep->dpe_pcache;
662 662 is_epoll = (dpep->dpe_flag & DP_ISEPOLLCOMPAT) != 0;
663 663 size = (is_epoll) ? sizeof (dvpoll_epollfd_t) : sizeof (pollfd_t);
664 664 mutex_exit(&dpep->dpe_lock);
665 665
666 666 if (!is_epoll && curproc->p_pid != pcp->pc_pid) {
667 667 if (pcp->pc_pid != -1) {
|
↓ open down ↓ |
667 lines elided |
↑ open up ↑ |
668 668 return (EACCES);
669 669 }
670 670
671 671 pcp->pc_pid = curproc->p_pid;
672 672 }
673 673
674 674 uiosize = uiop->uio_resid;
675 675 pollfdnum = uiosize / size;
676 676
677 677 /*
678 - * For epoll-enabled handles, restrict the allowed write size to 2.
679 - * This corresponds to an epoll_ctl(3C) performing an EPOLL_CTL_MOD
680 - * operation which is expanded into two operations (DEL and ADD).
681 - *
682 - * All other operations performed through epoll_ctl(3C) will consist of
683 - * a single entry.
684 - */
685 - if (is_epoll && pollfdnum > 2) {
686 - return (EINVAL);
687 - }
688 -
689 - /*
690 678 * We want to make sure that pollfdnum isn't large enough to DoS us,
691 679 * but we also don't want to grab p_lock unnecessarily -- so we
692 680 * perform the full check against our resource limits if and only if
693 681 * pollfdnum is larger than the known-to-be-sane value of UINT8_MAX.
694 682 */
695 683 if (pollfdnum > UINT8_MAX) {
696 684 mutex_enter(&curproc->p_lock);
697 685 if (pollfdnum >
698 686 (uint_t)rctl_enforced_value(rctlproc_legacy[RLIMIT_NOFILE],
699 687 curproc->p_rctls, curproc)) {
700 688 (void) rctl_action(rctlproc_legacy[RLIMIT_NOFILE],
701 689 curproc->p_rctls, curproc, RCA_SAFE);
702 690 mutex_exit(&curproc->p_lock);
703 691 return (EINVAL);
704 692 }
705 693 mutex_exit(&curproc->p_lock);
706 694 }
707 695
708 696 /*
709 697 * Copy in the pollfd array. Walk through the array and add
710 698 * each polled fd to the cached set.
711 699 */
712 700 pollfdp = kmem_alloc(uiosize, KM_SLEEP);
713 701 limit = (uintptr_t)pollfdp + (pollfdnum * size);
714 702
715 703 /*
716 704 * Although /dev/poll uses the write(2) interface to cache fds, it's
717 705 * not supposed to function as a seekable device. To prevent offset
718 706 * from growing and eventually exceed the maximum, reset the offset
719 707 * here for every call.
720 708 */
721 709 uiop->uio_loffset = 0;
722 710
723 711 /*
724 712 * Use uiocopy instead of uiomove when populating pollfdp, keeping
725 713 * uio_resid untouched for now. Write syscalls will translate EINTR
726 714 * into a success if they detect "successfully transfered" data via an
727 715 * updated uio_resid. Falsely suppressing such errors is disastrous.
728 716 */
729 717 if ((error = uiocopy((caddr_t)pollfdp, uiosize, UIO_WRITE, uiop,
730 718 ©size)) != 0) {
731 719 kmem_free(pollfdp, uiosize);
732 720 return (error);
733 721 }
734 722
735 723 /*
736 724 * We are about to enter the core portion of dpwrite(). Make sure this
737 725 * write has exclusive access in this portion of the code, i.e., no
|
↓ open down ↓ |
38 lines elided |
↑ open up ↑ |
738 726 * other writers in this code.
739 727 *
740 728 * Waiting for all readers to drop their references to the dpe is
741 729 * unecessary since the pollcache itself is protected by pc_lock.
742 730 */
743 731 mutex_enter(&dpep->dpe_lock);
744 732 dpep->dpe_writerwait++;
745 733 while ((dpep->dpe_flag & DP_WRITER_PRESENT) != 0) {
746 734 ASSERT(dpep->dpe_refcnt != 0);
747 735
748 - /*
749 - * The epoll API does not allow EINTR as a result when making
750 - * modifications to the set of polled fds. Given that write
751 - * activity is relatively quick and the size of accepted writes
752 - * is limited above to two entries, a signal-ignorant wait is
753 - * used here to avoid the EINTR.
754 - */
755 - if (is_epoll) {
756 - cv_wait(&dpep->dpe_cv, &dpep->dpe_lock);
757 - continue;
758 - }
759 -
760 - /*
761 - * Non-epoll writers to /dev/poll handles can tolerate EINTR.
762 - */
763 736 if (!cv_wait_sig_swap(&dpep->dpe_cv, &dpep->dpe_lock)) {
764 737 dpep->dpe_writerwait--;
765 738 mutex_exit(&dpep->dpe_lock);
766 739 kmem_free(pollfdp, uiosize);
767 740 return (EINTR);
768 741 }
769 742 }
770 743 dpep->dpe_writerwait--;
771 744 dpep->dpe_flag |= DP_WRITER_PRESENT;
772 745 dpep->dpe_refcnt++;
773 746
774 747 if (!is_epoll && (dpep->dpe_flag & DP_ISEPOLLCOMPAT) != 0) {
775 748 /*
776 749 * The epoll compat mode was enabled while we were waiting to
777 750 * establish write access. It is not safe to continue since
778 751 * state was prepared for non-epoll operation.
779 752 */
780 753 error = EBUSY;
781 754 goto bypass;
782 755 }
783 756 mutex_exit(&dpep->dpe_lock);
784 757
785 758 /*
786 759 * Since the dpwrite() may recursively walk an added /dev/poll handle,
787 760 * pollstate_enter() deadlock and loop detection must be used.
788 761 */
789 762 (void) pollstate_create();
790 763 VERIFY(pollstate_enter(pcp) == PSE_SUCCESS);
791 764
792 765 if (pcp->pc_bitmap == NULL) {
793 766 pcache_create(pcp, pollfdnum);
794 767 }
795 768 for (pfdp = pollfdp; (uintptr_t)pfdp < limit;
796 769 pfdp = (pollfd_t *)((uintptr_t)pfdp + size)) {
797 770 fd = pfdp->fd;
798 771 if ((uint_t)fd >= P_FINFO(curproc)->fi_nfiles) {
799 772 /*
800 773 * epoll semantics demand that we return EBADF if our
801 774 * specified fd is invalid.
802 775 */
803 776 if (is_epoll) {
804 777 error = EBADF;
805 778 break;
806 779 }
807 780
808 781 continue;
809 782 }
810 783
811 784 pdp = pcache_lookup_fd(pcp, fd);
812 785 if (pfdp->events != POLLREMOVE) {
813 786
814 787 fp = NULL;
815 788
816 789 if (pdp == NULL) {
817 790 /*
818 791 * If we're in epoll compatibility mode, check
819 792 * that the fd is valid before allocating
820 793 * anything for it; epoll semantics demand that
821 794 * we return EBADF if our specified fd is
822 795 * invalid.
823 796 */
824 797 if (is_epoll) {
825 798 if ((fp = getf(fd)) == NULL) {
826 799 error = EBADF;
827 800 break;
828 801 }
829 802 }
830 803
831 804 pdp = pcache_alloc_fd(0);
832 805 pdp->pd_fd = fd;
833 806 pdp->pd_pcache = pcp;
834 807 pcache_insert_fd(pcp, pdp, pollfdnum);
835 808 } else {
836 809 /*
837 810 * epoll semantics demand that we error out if
838 811 * a file descriptor is added twice, which we
839 812 * check (imperfectly) by checking if we both
840 813 * have the file descriptor cached and the
841 814 * file pointer that correponds to the file
842 815 * descriptor matches our cached value. If
843 816 * there is a pointer mismatch, the file
844 817 * descriptor was closed without being removed.
845 818 * The converse is clearly not true, however,
846 819 * so to narrow the window by which a spurious
847 820 * EEXIST may be returned, we also check if
848 821 * this fp has been added to an epoll control
849 822 * descriptor in the past; if it hasn't, we
850 823 * know that this is due to fp reuse -- it's
851 824 * not a true EEXIST case. (By performing this
852 825 * additional check, we limit the window of
853 826 * spurious EEXIST to situations where a single
854 827 * file descriptor is being used across two or
855 828 * more epoll control descriptors -- and even
856 829 * then, the file descriptor must be closed and
857 830 * reused in a relatively tight time span.)
858 831 */
859 832 if (is_epoll) {
860 833 if (pdp->pd_fp != NULL &&
861 834 (fp = getf(fd)) != NULL &&
862 835 fp == pdp->pd_fp &&
863 836 (fp->f_flag2 & FEPOLLED)) {
864 837 error = EEXIST;
865 838 releasef(fd);
866 839 break;
867 840 }
868 841
869 842 /*
870 843 * We have decided that the cached
871 844 * information was stale: it either
872 845 * didn't match, or the fp had never
873 846 * actually been epoll()'d on before.
874 847 * We need to now clear our pd_events
875 848 * to assure that we don't mistakenly
876 849 * operate on cached event disposition.
877 850 */
878 851 pdp->pd_events = 0;
879 852 }
880 853 }
881 854
882 855 if (is_epoll) {
883 856 epfdp = (dvpoll_epollfd_t *)pfdp;
884 857 pdp->pd_epolldata = epfdp->dpep_data;
885 858 }
886 859
887 860 ASSERT(pdp->pd_fd == fd);
888 861 ASSERT(pdp->pd_pcache == pcp);
889 862 if (fd >= pcp->pc_mapsize) {
890 863 mutex_exit(&pcp->pc_lock);
891 864 pcache_grow_map(pcp, fd);
892 865 mutex_enter(&pcp->pc_lock);
893 866 }
894 867 if (fd > pcp->pc_mapend) {
895 868 pcp->pc_mapend = fd;
896 869 }
897 870 if (fp == NULL && (fp = getf(fd)) == NULL) {
898 871 /*
899 872 * The fd is not valid. Since we can't pass
900 873 * this error back in the write() call, set
901 874 * the bit in bitmap to force DP_POLL ioctl
902 875 * to examine it.
903 876 */
904 877 BT_SET(pcp->pc_bitmap, fd);
905 878 pdp->pd_events |= pfdp->events;
906 879 continue;
907 880 }
908 881
909 882 /*
910 883 * To (greatly) reduce EEXIST false positives, we
911 884 * denote that this fp has been epoll()'d. We do this
912 885 * regardless of epoll compatibility mode, as the flag
913 886 * is harmless if not in epoll compatibility mode.
914 887 */
915 888 fp->f_flag2 |= FEPOLLED;
916 889
917 890 /*
918 891 * Don't do VOP_POLL for an already cached fd with
919 892 * same poll events.
920 893 */
921 894 if ((pdp->pd_events == pfdp->events) &&
922 895 (pdp->pd_fp == fp)) {
923 896 /*
924 897 * the events are already cached
925 898 */
926 899 releasef(fd);
927 900 continue;
928 901 }
929 902
930 903 /*
931 904 * do VOP_POLL and cache this poll fd.
932 905 */
933 906 /*
934 907 * XXX - pollrelock() logic needs to know which
935 908 * which pollcache lock to grab. It'd be a
936 909 * cleaner solution if we could pass pcp as
937 910 * an arguement in VOP_POLL interface instead
938 911 * of implicitly passing it using thread_t
939 912 * struct. On the other hand, changing VOP_POLL
940 913 * interface will require all driver/file system
941 914 * poll routine to change. May want to revisit
942 915 * the tradeoff later.
943 916 */
944 917 curthread->t_pollcache = pcp;
945 918 error = VOP_POLL(fp->f_vnode, pfdp->events, 0,
946 919 &pfdp->revents, &php, NULL);
947 920 curthread->t_pollcache = NULL;
948 921 /*
949 922 * We always set the bit when this fd is cached;
950 923 * this forces the first DP_POLL to poll this fd.
951 924 * Real performance gain comes from subsequent
952 925 * DP_POLL. We also attempt a pollhead_insert();
953 926 * if it's not possible, we'll do it in dpioctl().
954 927 */
955 928 BT_SET(pcp->pc_bitmap, fd);
956 929 if (error != 0) {
957 930 releasef(fd);
958 931 break;
959 932 }
960 933 pdp->pd_fp = fp;
961 934 pdp->pd_events |= pfdp->events;
962 935 if (php != NULL) {
963 936 if (pdp->pd_php == NULL) {
964 937 pollhead_insert(php, pdp);
965 938 pdp->pd_php = php;
966 939 } else {
967 940 if (pdp->pd_php != php) {
968 941 pollhead_delete(pdp->pd_php,
969 942 pdp);
970 943 pollhead_insert(php, pdp);
971 944 pdp->pd_php = php;
972 945 }
973 946 }
974 947 }
975 948 fds_added = B_TRUE;
976 949 releasef(fd);
977 950 } else {
978 951 if (pdp == NULL || pdp->pd_fp == NULL) {
979 952 if (is_epoll) {
980 953 /*
981 954 * As with the add case (above), epoll
982 955 * semantics demand that we error out
983 956 * in this case.
984 957 */
985 958 error = ENOENT;
986 959 break;
987 960 }
988 961
989 962 continue;
990 963 }
991 964 ASSERT(pdp->pd_fd == fd);
992 965 pdp->pd_fp = NULL;
993 966 pdp->pd_events = 0;
994 967 ASSERT(pdp->pd_thread == NULL);
995 968 if (pdp->pd_php != NULL) {
996 969 pollhead_delete(pdp->pd_php, pdp);
997 970 pdp->pd_php = NULL;
998 971 }
999 972 BT_CLEAR(pcp->pc_bitmap, fd);
1000 973 }
1001 974 }
1002 975 /*
1003 976 * Wake any pollcache waiters so they can check the new descriptors.
1004 977 *
1005 978 * Any fds added to an recursive-capable pollcache could themselves be
1006 979 * /dev/poll handles. To ensure that proper event propagation occurs,
1007 980 * parent pollcaches are woken too, so that they can create any needed
1008 981 * pollcache links.
1009 982 */
1010 983 if (fds_added) {
1011 984 cv_broadcast(&pcp->pc_cv);
1012 985 pcache_wake_parents(pcp);
1013 986 }
1014 987 pollstate_exit(pcp);
1015 988 mutex_enter(&dpep->dpe_lock);
1016 989 bypass:
1017 990 dpep->dpe_flag &= ~DP_WRITER_PRESENT;
1018 991 dpep->dpe_refcnt--;
1019 992 cv_broadcast(&dpep->dpe_cv);
1020 993 mutex_exit(&dpep->dpe_lock);
1021 994 kmem_free(pollfdp, uiosize);
1022 995 if (error == 0) {
1023 996 /*
1024 997 * The state of uio_resid is updated only after the pollcache
1025 998 * is successfully modified.
1026 999 */
1027 1000 uioskip(uiop, copysize);
1028 1001 }
1029 1002 return (error);
1030 1003 }
1031 1004
1032 1005 #define DP_SIGMASK_RESTORE(ksetp) { \
1033 1006 if (ksetp != NULL) { \
1034 1007 mutex_enter(&p->p_lock); \
1035 1008 if (lwp->lwp_cursig == 0) { \
1036 1009 t->t_hold = lwp->lwp_sigoldmask; \
1037 1010 t->t_flag &= ~T_TOMASK; \
1038 1011 } \
1039 1012 mutex_exit(&p->p_lock); \
1040 1013 } \
1041 1014 }
1042 1015
1043 1016 /*ARGSUSED*/
1044 1017 static int
1045 1018 dpioctl(dev_t dev, int cmd, intptr_t arg, int mode, cred_t *credp, int *rvalp)
1046 1019 {
1047 1020 minor_t minor;
1048 1021 dp_entry_t *dpep;
1049 1022 pollcache_t *pcp;
1050 1023 hrtime_t now;
1051 1024 int error = 0;
1052 1025 boolean_t is_epoll;
1053 1026 STRUCT_DECL(dvpoll, dvpoll);
1054 1027
1055 1028 if (cmd == DP_POLL || cmd == DP_PPOLL) {
1056 1029 /* do this now, before we sleep on DP_WRITER_PRESENT */
1057 1030 now = gethrtime();
1058 1031 }
1059 1032
1060 1033 minor = getminor(dev);
1061 1034 mutex_enter(&devpoll_lock);
1062 1035 ASSERT(minor < dptblsize);
1063 1036 dpep = devpolltbl[minor];
1064 1037 mutex_exit(&devpoll_lock);
1065 1038 ASSERT(dpep != NULL);
1066 1039 pcp = dpep->dpe_pcache;
1067 1040
1068 1041 mutex_enter(&dpep->dpe_lock);
1069 1042 is_epoll = (dpep->dpe_flag & DP_ISEPOLLCOMPAT) != 0;
1070 1043
1071 1044 if (cmd == DP_EPOLLCOMPAT) {
1072 1045 if (dpep->dpe_refcnt != 0) {
1073 1046 /*
1074 1047 * We can't turn on epoll compatibility while there
1075 1048 * are outstanding operations.
1076 1049 */
1077 1050 mutex_exit(&dpep->dpe_lock);
1078 1051 return (EBUSY);
1079 1052 }
1080 1053
1081 1054 /*
1082 1055 * epoll compatibility is a one-way street: there's no way
1083 1056 * to turn it off for a particular open.
1084 1057 */
1085 1058 dpep->dpe_flag |= DP_ISEPOLLCOMPAT;
1086 1059 mutex_exit(&dpep->dpe_lock);
1087 1060
1088 1061 return (0);
1089 1062 }
1090 1063
1091 1064 if (!is_epoll && curproc->p_pid != pcp->pc_pid) {
1092 1065 if (pcp->pc_pid != -1) {
1093 1066 mutex_exit(&dpep->dpe_lock);
1094 1067 return (EACCES);
1095 1068 }
1096 1069
1097 1070 pcp->pc_pid = curproc->p_pid;
1098 1071 }
1099 1072
1100 1073 /* Wait until all writers have cleared the handle before continuing */
1101 1074 while ((dpep->dpe_flag & DP_WRITER_PRESENT) != 0 ||
1102 1075 (dpep->dpe_writerwait != 0)) {
1103 1076 if (!cv_wait_sig_swap(&dpep->dpe_cv, &dpep->dpe_lock)) {
1104 1077 mutex_exit(&dpep->dpe_lock);
1105 1078 return (EINTR);
1106 1079 }
1107 1080 }
1108 1081 dpep->dpe_refcnt++;
1109 1082 mutex_exit(&dpep->dpe_lock);
1110 1083
1111 1084 switch (cmd) {
1112 1085 case DP_POLL:
1113 1086 case DP_PPOLL:
1114 1087 {
1115 1088 pollstate_t *ps;
1116 1089 nfds_t nfds;
1117 1090 int fdcnt = 0;
1118 1091 size_t size, fdsize, dpsize;
1119 1092 hrtime_t deadline = 0;
1120 1093 k_sigset_t *ksetp = NULL;
1121 1094 k_sigset_t kset;
1122 1095 sigset_t set;
1123 1096 kthread_t *t = curthread;
1124 1097 klwp_t *lwp = ttolwp(t);
1125 1098 struct proc *p = ttoproc(curthread);
1126 1099
1127 1100 STRUCT_INIT(dvpoll, mode);
1128 1101
1129 1102 /*
1130 1103 * The dp_setp member is only required/consumed for DP_PPOLL,
1131 1104 * which otherwise uses the same structure as DP_POLL.
1132 1105 */
1133 1106 if (cmd == DP_POLL) {
1134 1107 dpsize = (uintptr_t)STRUCT_FADDR(dvpoll, dp_setp) -
1135 1108 (uintptr_t)STRUCT_FADDR(dvpoll, dp_fds);
1136 1109 } else {
1137 1110 ASSERT(cmd == DP_PPOLL);
1138 1111 dpsize = STRUCT_SIZE(dvpoll);
1139 1112 }
1140 1113
1141 1114 if ((mode & FKIOCTL) != 0) {
1142 1115 /* Kernel-internal ioctl call */
1143 1116 bcopy((caddr_t)arg, STRUCT_BUF(dvpoll), dpsize);
1144 1117 error = 0;
1145 1118 } else {
1146 1119 error = copyin((caddr_t)arg, STRUCT_BUF(dvpoll),
1147 1120 dpsize);
1148 1121 }
1149 1122
1150 1123 if (error) {
1151 1124 DP_REFRELE(dpep);
1152 1125 return (EFAULT);
1153 1126 }
1154 1127
1155 1128 deadline = STRUCT_FGET(dvpoll, dp_timeout);
1156 1129 if (deadline > 0) {
1157 1130 /*
1158 1131 * Convert the deadline from relative milliseconds
1159 1132 * to absolute nanoseconds. They must wait for at
1160 1133 * least a tick.
1161 1134 */
1162 1135 deadline = MSEC2NSEC(deadline);
1163 1136 deadline = MAX(deadline, nsec_per_tick);
1164 1137 deadline += now;
1165 1138 }
1166 1139
1167 1140 if (cmd == DP_PPOLL) {
1168 1141 void *setp = STRUCT_FGETP(dvpoll, dp_setp);
1169 1142
1170 1143 if (setp != NULL) {
1171 1144 if ((mode & FKIOCTL) != 0) {
1172 1145 /* Use the signal set directly */
1173 1146 ksetp = (k_sigset_t *)setp;
1174 1147 } else {
1175 1148 if (copyin(setp, &set, sizeof (set))) {
1176 1149 DP_REFRELE(dpep);
1177 1150 return (EFAULT);
1178 1151 }
1179 1152 sigutok(&set, &kset);
1180 1153 ksetp = &kset;
1181 1154 }
1182 1155
1183 1156 mutex_enter(&p->p_lock);
1184 1157 schedctl_finish_sigblock(t);
1185 1158 lwp->lwp_sigoldmask = t->t_hold;
1186 1159 t->t_hold = *ksetp;
1187 1160 t->t_flag |= T_TOMASK;
1188 1161
1189 1162 /*
1190 1163 * Like ppoll() with a non-NULL sigset, we'll
1191 1164 * call cv_reltimedwait_sig() just to check for
1192 1165 * signals. This call will return immediately
1193 1166 * with either 0 (signalled) or -1 (no signal).
1194 1167 * There are some conditions whereby we can
1195 1168 * get 0 from cv_reltimedwait_sig() without
1196 1169 * a true signal (e.g., a directed stop), so
1197 1170 * we restore our signal mask in the unlikely
1198 1171 * event that lwp_cursig is 0.
1199 1172 */
1200 1173 if (!cv_reltimedwait_sig(&t->t_delay_cv,
1201 1174 &p->p_lock, 0, TR_CLOCK_TICK)) {
1202 1175 if (lwp->lwp_cursig == 0) {
1203 1176 t->t_hold = lwp->lwp_sigoldmask;
1204 1177 t->t_flag &= ~T_TOMASK;
1205 1178 }
1206 1179
1207 1180 mutex_exit(&p->p_lock);
1208 1181
1209 1182 DP_REFRELE(dpep);
1210 1183 return (EINTR);
1211 1184 }
1212 1185
1213 1186 mutex_exit(&p->p_lock);
1214 1187 }
1215 1188 }
1216 1189
1217 1190 if ((nfds = STRUCT_FGET(dvpoll, dp_nfds)) == 0) {
1218 1191 /*
1219 1192 * We are just using DP_POLL to sleep, so
1220 1193 * we don't any of the devpoll apparatus.
1221 1194 * Do not check for signals if we have a zero timeout.
1222 1195 */
1223 1196 DP_REFRELE(dpep);
1224 1197 if (deadline == 0) {
1225 1198 DP_SIGMASK_RESTORE(ksetp);
1226 1199 return (0);
1227 1200 }
1228 1201
1229 1202 mutex_enter(&curthread->t_delay_lock);
1230 1203 while ((error =
1231 1204 cv_timedwait_sig_hrtime(&curthread->t_delay_cv,
1232 1205 &curthread->t_delay_lock, deadline)) > 0)
1233 1206 continue;
1234 1207 mutex_exit(&curthread->t_delay_lock);
1235 1208
1236 1209 DP_SIGMASK_RESTORE(ksetp);
1237 1210
1238 1211 return (error == 0 ? EINTR : 0);
1239 1212 }
1240 1213
1241 1214 if (is_epoll) {
1242 1215 size = nfds * (fdsize = sizeof (epoll_event_t));
1243 1216 } else {
1244 1217 size = nfds * (fdsize = sizeof (pollfd_t));
1245 1218 }
1246 1219
1247 1220 /*
1248 1221 * XXX It would be nice not to have to alloc each time, but it
1249 1222 * requires another per thread structure hook. This can be
1250 1223 * implemented later if data suggests that it's necessary.
1251 1224 */
1252 1225 ps = pollstate_create();
1253 1226
1254 1227 if (ps->ps_dpbufsize < size) {
1255 1228 /*
1256 1229 * If nfds is larger than twice the current maximum
1257 1230 * open file count, we'll silently clamp it. This
1258 1231 * only limits our exposure to allocating an
1259 1232 * inordinate amount of kernel memory; it doesn't
1260 1233 * otherwise affect the semantics. (We have this
1261 1234 * check at twice the maximum instead of merely the
1262 1235 * maximum because some applications pass an nfds that
1263 1236 * is only slightly larger than their limit.)
1264 1237 */
1265 1238 mutex_enter(&p->p_lock);
1266 1239 if ((nfds >> 1) > p->p_fno_ctl) {
1267 1240 nfds = p->p_fno_ctl;
1268 1241 size = nfds * fdsize;
1269 1242 }
1270 1243 mutex_exit(&p->p_lock);
1271 1244
1272 1245 if (ps->ps_dpbufsize < size) {
1273 1246 kmem_free(ps->ps_dpbuf, ps->ps_dpbufsize);
1274 1247 ps->ps_dpbuf = kmem_zalloc(size, KM_SLEEP);
1275 1248 ps->ps_dpbufsize = size;
1276 1249 }
1277 1250 }
1278 1251
1279 1252 VERIFY(pollstate_enter(pcp) == PSE_SUCCESS);
1280 1253 for (;;) {
1281 1254 pcp->pc_flag &= ~PC_POLLWAKE;
1282 1255
1283 1256 /*
1284 1257 * Mark all child pcachelinks as stale.
1285 1258 * Those which are still part of the tree will be
1286 1259 * marked as valid during the poll.
1287 1260 */
1288 1261 pcachelink_mark_stale(pcp);
1289 1262
1290 1263 error = dp_pcache_poll(dpep, ps->ps_dpbuf,
1291 1264 pcp, nfds, &fdcnt);
1292 1265 if (fdcnt > 0 || error != 0)
1293 1266 break;
1294 1267
1295 1268 /* Purge still-stale child pcachelinks */
1296 1269 pcachelink_purge_stale(pcp);
1297 1270
1298 1271 /*
1299 1272 * A pollwake has happened since we polled cache.
1300 1273 */
1301 1274 if (pcp->pc_flag & PC_POLLWAKE)
1302 1275 continue;
1303 1276
1304 1277 /*
1305 1278 * Sleep until we are notified, signaled, or timed out.
1306 1279 */
1307 1280 if (deadline == 0) {
1308 1281 /* immediate timeout; do not check signals */
1309 1282 break;
1310 1283 }
1311 1284
1312 1285 error = cv_timedwait_sig_hrtime(&pcp->pc_cv,
1313 1286 &pcp->pc_lock, deadline);
1314 1287
1315 1288 /*
1316 1289 * If we were awakened by a signal or timeout then
1317 1290 * break the loop, else poll again.
1318 1291 */
1319 1292 if (error <= 0) {
1320 1293 error = (error == 0) ? EINTR : 0;
1321 1294 break;
1322 1295 } else {
1323 1296 error = 0;
1324 1297 }
1325 1298 }
1326 1299 pollstate_exit(pcp);
1327 1300
1328 1301 DP_SIGMASK_RESTORE(ksetp);
1329 1302
1330 1303 if (error == 0 && fdcnt > 0) {
1331 1304 /*
1332 1305 * It should be noted that FKIOCTL does not influence
1333 1306 * the copyout (vs bcopy) of dp_fds at this time.
1334 1307 */
1335 1308 if (copyout(ps->ps_dpbuf,
1336 1309 STRUCT_FGETP(dvpoll, dp_fds), fdcnt * fdsize)) {
1337 1310 DP_REFRELE(dpep);
1338 1311 return (EFAULT);
1339 1312 }
1340 1313 *rvalp = fdcnt;
1341 1314 }
1342 1315 break;
1343 1316 }
1344 1317
1345 1318 case DP_ISPOLLED:
1346 1319 {
1347 1320 pollfd_t pollfd;
1348 1321 polldat_t *pdp;
1349 1322
1350 1323 STRUCT_INIT(dvpoll, mode);
1351 1324 error = copyin((caddr_t)arg, &pollfd, sizeof (pollfd_t));
1352 1325 if (error) {
1353 1326 DP_REFRELE(dpep);
1354 1327 return (EFAULT);
1355 1328 }
1356 1329 mutex_enter(&pcp->pc_lock);
1357 1330 if (pcp->pc_hash == NULL) {
1358 1331 /*
1359 1332 * No Need to search because no poll fd
1360 1333 * has been cached.
1361 1334 */
1362 1335 mutex_exit(&pcp->pc_lock);
1363 1336 DP_REFRELE(dpep);
1364 1337 return (0);
1365 1338 }
1366 1339 if (pollfd.fd < 0) {
1367 1340 mutex_exit(&pcp->pc_lock);
1368 1341 break;
1369 1342 }
1370 1343 pdp = pcache_lookup_fd(pcp, pollfd.fd);
1371 1344 if ((pdp != NULL) && (pdp->pd_fd == pollfd.fd) &&
1372 1345 (pdp->pd_fp != NULL)) {
1373 1346 pollfd.revents = pdp->pd_events;
1374 1347 if (copyout(&pollfd, (caddr_t)arg, sizeof (pollfd_t))) {
1375 1348 mutex_exit(&pcp->pc_lock);
1376 1349 DP_REFRELE(dpep);
1377 1350 return (EFAULT);
1378 1351 }
1379 1352 *rvalp = 1;
1380 1353 }
1381 1354 mutex_exit(&pcp->pc_lock);
1382 1355 break;
1383 1356 }
1384 1357
1385 1358 default:
1386 1359 DP_REFRELE(dpep);
1387 1360 return (EINVAL);
1388 1361 }
1389 1362 DP_REFRELE(dpep);
1390 1363 return (error);
1391 1364 }
1392 1365
1393 1366 /*
1394 1367 * Overview of Recursive Polling
1395 1368 *
1396 1369 * It is possible for /dev/poll to poll for events on file descriptors which
1397 1370 * themselves are /dev/poll handles. Pending events in the child handle are
1398 1371 * represented as readable data via the POLLIN flag. To limit surface area,
1399 1372 * this recursion is presently allowed on only /dev/poll handles which have
1400 1373 * been placed in epoll mode via the DP_EPOLLCOMPAT ioctl. Recursion depth is
1401 1374 * limited to 5 in order to be consistent with Linux epoll.
1402 1375 *
1403 1376 * Extending dppoll() for VOP_POLL:
1404 1377 *
1405 1378 * The recursive /dev/poll implementation begins by extending dppoll() to
1406 1379 * report when resources contained in the pollcache have relevant event state.
1407 1380 * At the highest level, it means calling dp_pcache_poll() so it indicates if
1408 1381 * fd events are present without consuming them or altering the pollcache
1409 1382 * bitmap. This ensures that a subsequent DP_POLL operation on the bitmap will
1410 1383 * yield the initiating event. Additionally, the VOP_POLL should return in
1411 1384 * such a way that dp_pcache_poll() does not clear the parent bitmap entry
1412 1385 * which corresponds to the child /dev/poll fd. This means that child
1413 1386 * pollcaches will be checked during every poll which facilitates wake-up
1414 1387 * behavior detailed below.
1415 1388 *
1416 1389 * Pollcache Links and Wake Events:
1417 1390 *
1418 1391 * Recursive /dev/poll avoids complicated pollcache locking constraints during
1419 1392 * pollwakeup events by eschewing the traditional pollhead mechanism in favor
1420 1393 * of a different approach. For each pollcache at the root of a recursive
1421 1394 * /dev/poll "tree", pcachelink_t structures are established to all child
1422 1395 * /dev/poll pollcaches. During pollnotify() in a child pollcache, the
1423 1396 * linked list of pcachelink_t entries is walked, where those marked as valid
1424 1397 * incur a cv_broadcast to their parent pollcache. Most notably, these
1425 1398 * pcachelink_t cv wakeups are performed without acquiring pc_lock on the
1426 1399 * parent pollcache (which would require careful deadlock avoidance). This
1427 1400 * still allows the woken poll on the parent to discover the pertinent events
1428 1401 * due to the fact that bitmap entires for the child pollcache are always
1429 1402 * maintained by the dppoll() logic above.
1430 1403 *
1431 1404 * Depth Limiting and Loop Prevention:
1432 1405 *
1433 1406 * As each pollcache is encountered (either via DP_POLL or dppoll()), depth and
1434 1407 * loop constraints are enforced via pollstate_enter(). The pollcache_t
1435 1408 * pointer is compared against any existing entries in ps_pc_stack and is added
1436 1409 * to the end if no match (and therefore loop) is found. Once poll operations
1437 1410 * for a given pollcache_t are complete, pollstate_exit() clears the pointer
1438 1411 * from the list. The pollstate_enter() and pollstate_exit() functions are
1439 1412 * responsible for acquiring and releasing pc_lock, respectively.
1440 1413 *
1441 1414 * Deadlock Safety:
1442 1415 *
1443 1416 * Descending through a tree of recursive /dev/poll handles involves the tricky
1444 1417 * business of sequentially entering multiple pollcache locks. This tree
1445 1418 * topology cannot define a lock acquisition order in such a way that it is
1446 1419 * immune to deadlocks between threads. The pollstate_enter() and
1447 1420 * pollstate_exit() functions provide an interface for recursive /dev/poll
1448 1421 * operations to safely lock pollcaches while failing gracefully in the face of
1449 1422 * deadlocking topologies. (See pollstate_contend() for more detail about how
1450 1423 * deadlocks are detected and resolved.)
1451 1424 */
1452 1425
1453 1426 /*ARGSUSED*/
1454 1427 static int
1455 1428 dppoll(dev_t dev, short events, int anyyet, short *reventsp,
1456 1429 struct pollhead **phpp)
1457 1430 {
1458 1431 minor_t minor;
1459 1432 dp_entry_t *dpep;
1460 1433 pollcache_t *pcp;
1461 1434 int res, rc = 0;
1462 1435
1463 1436 minor = getminor(dev);
1464 1437 mutex_enter(&devpoll_lock);
1465 1438 ASSERT(minor < dptblsize);
1466 1439 dpep = devpolltbl[minor];
1467 1440 ASSERT(dpep != NULL);
1468 1441 mutex_exit(&devpoll_lock);
1469 1442
1470 1443 mutex_enter(&dpep->dpe_lock);
1471 1444 if ((dpep->dpe_flag & DP_ISEPOLLCOMPAT) == 0) {
1472 1445 /* Poll recursion is not yet supported for non-epoll handles */
1473 1446 *reventsp = POLLERR;
1474 1447 mutex_exit(&dpep->dpe_lock);
1475 1448 return (0);
1476 1449 } else {
1477 1450 dpep->dpe_refcnt++;
1478 1451 pcp = dpep->dpe_pcache;
1479 1452 mutex_exit(&dpep->dpe_lock);
1480 1453 }
1481 1454
1482 1455 res = pollstate_enter(pcp);
1483 1456 if (res == PSE_SUCCESS) {
1484 1457 nfds_t nfds = 1;
1485 1458 int fdcnt = 0;
1486 1459 pollstate_t *ps = curthread->t_pollstate;
1487 1460
1488 1461 rc = dp_pcache_poll(dpep, NULL, pcp, nfds, &fdcnt);
1489 1462 if (rc == 0) {
1490 1463 *reventsp = (fdcnt > 0) ? POLLIN : 0;
1491 1464 }
1492 1465 pcachelink_assoc(pcp, ps->ps_pc_stack[0]);
1493 1466 pollstate_exit(pcp);
1494 1467 } else {
1495 1468 switch (res) {
1496 1469 case PSE_FAIL_DEPTH:
1497 1470 rc = EINVAL;
1498 1471 break;
1499 1472 case PSE_FAIL_LOOP:
1500 1473 case PSE_FAIL_DEADLOCK:
1501 1474 rc = ELOOP;
1502 1475 break;
1503 1476 default:
1504 1477 /*
1505 1478 * If anything else has gone awry, such as being polled
1506 1479 * from an unexpected context, fall back to the
1507 1480 * recursion-intolerant response.
1508 1481 */
1509 1482 *reventsp = POLLERR;
1510 1483 rc = 0;
1511 1484 break;
1512 1485 }
1513 1486 }
1514 1487
1515 1488 DP_REFRELE(dpep);
1516 1489 return (rc);
1517 1490 }
1518 1491
1519 1492 /*
1520 1493 * devpoll close should do enough clean up before the pollcache is deleted,
1521 1494 * i.e., it should ensure no one still references the pollcache later.
1522 1495 * There is no "permission" check in here. Any process having the last
1523 1496 * reference of this /dev/poll fd can close.
1524 1497 */
1525 1498 /*ARGSUSED*/
1526 1499 static int
1527 1500 dpclose(dev_t dev, int flag, int otyp, cred_t *credp)
1528 1501 {
1529 1502 minor_t minor;
1530 1503 dp_entry_t *dpep;
1531 1504 pollcache_t *pcp;
1532 1505 int i;
1533 1506 polldat_t **hashtbl;
1534 1507 polldat_t *pdp;
1535 1508
1536 1509 minor = getminor(dev);
1537 1510
1538 1511 mutex_enter(&devpoll_lock);
1539 1512 dpep = devpolltbl[minor];
1540 1513 ASSERT(dpep != NULL);
1541 1514 devpolltbl[minor] = NULL;
1542 1515 mutex_exit(&devpoll_lock);
1543 1516 pcp = dpep->dpe_pcache;
1544 1517 ASSERT(pcp != NULL);
1545 1518 /*
1546 1519 * At this point, no other lwp can access this pollcache via the
1547 1520 * /dev/poll fd. This pollcache is going away, so do the clean
1548 1521 * up without the pc_lock.
1549 1522 */
1550 1523 hashtbl = pcp->pc_hash;
1551 1524 for (i = 0; i < pcp->pc_hashsize; i++) {
1552 1525 for (pdp = hashtbl[i]; pdp; pdp = pdp->pd_hashnext) {
1553 1526 if (pdp->pd_php != NULL) {
1554 1527 pollhead_delete(pdp->pd_php, pdp);
1555 1528 pdp->pd_php = NULL;
1556 1529 pdp->pd_fp = NULL;
1557 1530 }
1558 1531 }
1559 1532 }
1560 1533 /*
1561 1534 * pollwakeup() may still interact with this pollcache. Wait until
1562 1535 * it is done.
1563 1536 */
1564 1537 mutex_enter(&pcp->pc_no_exit);
1565 1538 ASSERT(pcp->pc_busy >= 0);
1566 1539 while (pcp->pc_busy > 0)
1567 1540 cv_wait(&pcp->pc_busy_cv, &pcp->pc_no_exit);
1568 1541 mutex_exit(&pcp->pc_no_exit);
1569 1542
1570 1543 /* Clean up any pollcache links created via recursive /dev/poll */
1571 1544 if (pcp->pc_parents != NULL || pcp->pc_children != NULL) {
1572 1545 /*
1573 1546 * Because of the locking rules for pcachelink manipulation,
1574 1547 * acquring pc_lock is required for this step.
1575 1548 */
1576 1549 mutex_enter(&pcp->pc_lock);
1577 1550 pcachelink_purge_all(pcp);
1578 1551 mutex_exit(&pcp->pc_lock);
1579 1552 }
1580 1553
1581 1554 pcache_destroy(pcp);
1582 1555 ASSERT(dpep->dpe_refcnt == 0);
1583 1556 kmem_free(dpep, sizeof (dp_entry_t));
1584 1557 return (0);
1585 1558 }
1586 1559
1587 1560 static void
1588 1561 pcachelink_locked_rele(pcachelink_t *pl)
1589 1562 {
1590 1563 ASSERT(MUTEX_HELD(&pl->pcl_lock));
1591 1564 VERIFY(pl->pcl_refcnt >= 1);
1592 1565
1593 1566 pl->pcl_refcnt--;
1594 1567 if (pl->pcl_refcnt == 0) {
1595 1568 VERIFY(pl->pcl_state == PCL_INVALID);
1596 1569 ASSERT(pl->pcl_parent_pc == NULL);
1597 1570 ASSERT(pl->pcl_child_pc == NULL);
1598 1571 ASSERT(pl->pcl_parent_next == NULL);
1599 1572 ASSERT(pl->pcl_child_next == NULL);
1600 1573
1601 1574 pl->pcl_state = PCL_FREE;
1602 1575 mutex_destroy(&pl->pcl_lock);
1603 1576 kmem_free(pl, sizeof (pcachelink_t));
1604 1577 } else {
1605 1578 mutex_exit(&pl->pcl_lock);
1606 1579 }
1607 1580 }
1608 1581
1609 1582 /*
1610 1583 * Associate parent and child pollcaches via a pcachelink_t. If an existing
1611 1584 * link (stale or valid) between the two is found, it will be reused. If a
1612 1585 * suitable link is not found for reuse, a new one will be allocated.
1613 1586 */
1614 1587 static void
1615 1588 pcachelink_assoc(pollcache_t *child, pollcache_t *parent)
1616 1589 {
1617 1590 pcachelink_t *pl, **plpn;
1618 1591
1619 1592 ASSERT(MUTEX_HELD(&child->pc_lock));
1620 1593 ASSERT(MUTEX_HELD(&parent->pc_lock));
1621 1594
1622 1595 /* Search for an existing link we can reuse. */
1623 1596 plpn = &child->pc_parents;
1624 1597 for (pl = child->pc_parents; pl != NULL; pl = *plpn) {
1625 1598 mutex_enter(&pl->pcl_lock);
1626 1599 if (pl->pcl_state == PCL_INVALID) {
1627 1600 /* Clean any invalid links while walking the list */
1628 1601 *plpn = pl->pcl_parent_next;
1629 1602 pl->pcl_child_pc = NULL;
1630 1603 pl->pcl_parent_next = NULL;
1631 1604 pcachelink_locked_rele(pl);
1632 1605 } else if (pl->pcl_parent_pc == parent) {
1633 1606 /* Successfully found parent link */
1634 1607 ASSERT(pl->pcl_state == PCL_VALID ||
1635 1608 pl->pcl_state == PCL_STALE);
1636 1609 pl->pcl_state = PCL_VALID;
1637 1610 mutex_exit(&pl->pcl_lock);
1638 1611 return;
1639 1612 } else {
1640 1613 plpn = &pl->pcl_parent_next;
1641 1614 mutex_exit(&pl->pcl_lock);
1642 1615 }
1643 1616 }
1644 1617
1645 1618 /* No existing link to the parent was found. Create a fresh one. */
1646 1619 pl = kmem_zalloc(sizeof (pcachelink_t), KM_SLEEP);
1647 1620 mutex_init(&pl->pcl_lock, NULL, MUTEX_DEFAULT, NULL);
1648 1621
1649 1622 pl->pcl_parent_pc = parent;
1650 1623 pl->pcl_child_next = parent->pc_children;
1651 1624 parent->pc_children = pl;
1652 1625 pl->pcl_refcnt++;
1653 1626
1654 1627 pl->pcl_child_pc = child;
1655 1628 pl->pcl_parent_next = child->pc_parents;
1656 1629 child->pc_parents = pl;
1657 1630 pl->pcl_refcnt++;
1658 1631
1659 1632 pl->pcl_state = PCL_VALID;
1660 1633 }
1661 1634
1662 1635 /*
1663 1636 * Mark all child links in a pollcache as stale. Any invalid child links found
1664 1637 * during iteration are purged.
1665 1638 */
1666 1639 static void
1667 1640 pcachelink_mark_stale(pollcache_t *pcp)
1668 1641 {
1669 1642 pcachelink_t *pl, **plpn;
1670 1643
1671 1644 ASSERT(MUTEX_HELD(&pcp->pc_lock));
1672 1645
1673 1646 plpn = &pcp->pc_children;
1674 1647 for (pl = pcp->pc_children; pl != NULL; pl = *plpn) {
1675 1648 mutex_enter(&pl->pcl_lock);
1676 1649 if (pl->pcl_state == PCL_INVALID) {
1677 1650 /*
1678 1651 * Remove any invalid links while we are going to the
1679 1652 * trouble of walking the list.
1680 1653 */
1681 1654 *plpn = pl->pcl_child_next;
1682 1655 pl->pcl_parent_pc = NULL;
1683 1656 pl->pcl_child_next = NULL;
1684 1657 pcachelink_locked_rele(pl);
1685 1658 } else {
1686 1659 pl->pcl_state = PCL_STALE;
1687 1660 plpn = &pl->pcl_child_next;
1688 1661 mutex_exit(&pl->pcl_lock);
1689 1662 }
1690 1663 }
1691 1664 }
1692 1665
1693 1666 /*
1694 1667 * Purge all stale (or invalid) child links from a pollcache.
1695 1668 */
1696 1669 static void
1697 1670 pcachelink_purge_stale(pollcache_t *pcp)
1698 1671 {
1699 1672 pcachelink_t *pl, **plpn;
1700 1673
1701 1674 ASSERT(MUTEX_HELD(&pcp->pc_lock));
1702 1675
1703 1676 plpn = &pcp->pc_children;
1704 1677 for (pl = pcp->pc_children; pl != NULL; pl = *plpn) {
1705 1678 mutex_enter(&pl->pcl_lock);
1706 1679 switch (pl->pcl_state) {
1707 1680 case PCL_STALE:
1708 1681 pl->pcl_state = PCL_INVALID;
1709 1682 /* FALLTHROUGH */
1710 1683 case PCL_INVALID:
1711 1684 *plpn = pl->pcl_child_next;
1712 1685 pl->pcl_parent_pc = NULL;
1713 1686 pl->pcl_child_next = NULL;
1714 1687 pcachelink_locked_rele(pl);
1715 1688 break;
1716 1689 default:
1717 1690 plpn = &pl->pcl_child_next;
1718 1691 mutex_exit(&pl->pcl_lock);
1719 1692 }
1720 1693 }
1721 1694 }
1722 1695
1723 1696 /*
1724 1697 * Purge all child and parent links from a pollcache, regardless of status.
1725 1698 */
1726 1699 static void
1727 1700 pcachelink_purge_all(pollcache_t *pcp)
1728 1701 {
1729 1702 pcachelink_t *pl, **plpn;
1730 1703
1731 1704 ASSERT(MUTEX_HELD(&pcp->pc_lock));
1732 1705
1733 1706 plpn = &pcp->pc_parents;
1734 1707 for (pl = pcp->pc_parents; pl != NULL; pl = *plpn) {
1735 1708 mutex_enter(&pl->pcl_lock);
1736 1709 pl->pcl_state = PCL_INVALID;
1737 1710 *plpn = pl->pcl_parent_next;
1738 1711 pl->pcl_child_pc = NULL;
1739 1712 pl->pcl_parent_next = NULL;
1740 1713 pcachelink_locked_rele(pl);
1741 1714 }
1742 1715
1743 1716 plpn = &pcp->pc_children;
1744 1717 for (pl = pcp->pc_children; pl != NULL; pl = *plpn) {
1745 1718 mutex_enter(&pl->pcl_lock);
1746 1719 pl->pcl_state = PCL_INVALID;
1747 1720 *plpn = pl->pcl_child_next;
1748 1721 pl->pcl_parent_pc = NULL;
1749 1722 pl->pcl_child_next = NULL;
1750 1723 pcachelink_locked_rele(pl);
1751 1724 }
1752 1725
1753 1726 ASSERT(pcp->pc_parents == NULL);
1754 1727 ASSERT(pcp->pc_children == NULL);
1755 1728 }
|
↓ open down ↓ |
983 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX