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