1 /* $NetBSD: sys_select.c,v 1.57 2021/12/10 20:36:04 andvar Exp $ */ 2 3 /*- 4 * Copyright (c) 2007, 2008, 2009, 2010, 2019, 2020 The NetBSD Foundation, Inc. 5 * All rights reserved. 6 * 7 * This code is derived from software contributed to The NetBSD Foundation 8 * by Andrew Doran and Mindaugas Rasiukevicius. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 19 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 20 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 21 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 22 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 23 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 24 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 25 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 26 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 27 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 28 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 29 * POSSIBILITY OF SUCH DAMAGE. 30 */ 31 32 /* 33 * Copyright (c) 1982, 1986, 1989, 1993 34 * The Regents of the University of California. All rights reserved. 35 * (c) UNIX System Laboratories, Inc. 36 * All or some portions of this file are derived from material licensed 37 * to the University of California by American Telephone and Telegraph 38 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 39 * the permission of UNIX System Laboratories, Inc. 40 * 41 * Redistribution and use in source and binary forms, with or without 42 * modification, are permitted provided that the following conditions 43 * are met: 44 * 1. Redistributions of source code must retain the above copyright 45 * notice, this list of conditions and the following disclaimer. 46 * 2. Redistributions in binary form must reproduce the above copyright 47 * notice, this list of conditions and the following disclaimer in the 48 * documentation and/or other materials provided with the distribution. 49 * 3. Neither the name of the University nor the names of its contributors 50 * may be used to endorse or promote products derived from this software 51 * without specific prior written permission. 52 * 53 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 54 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 55 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 56 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 57 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 58 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 59 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 60 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 61 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 62 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 63 * SUCH DAMAGE. 64 * 65 * @(#)sys_generic.c 8.9 (Berkeley) 2/14/95 66 */ 67 68 /* 69 * System calls of synchronous I/O multiplexing subsystem. 70 * 71 * Locking 72 * 73 * Two locks are used: <object-lock> and selcluster_t::sc_lock. 74 * 75 * The <object-lock> might be a device driver or another subsystem, e.g. 76 * socket or pipe. This lock is not exported, and thus invisible to this 77 * subsystem. Mainly, synchronisation between selrecord() and selnotify() 78 * routines depends on this lock, as it will be described in the comments. 79 * 80 * Lock order 81 * 82 * <object-lock> -> 83 * selcluster_t::sc_lock 84 */ 85 86 #include <sys/cdefs.h> 87 __KERNEL_RCSID(0, "$NetBSD: sys_select.c,v 1.57 2021/12/10 20:36:04 andvar Exp $"); 88 89 #include <sys/param.h> 90 #include <sys/systm.h> 91 #include <sys/filedesc.h> 92 #include <sys/file.h> 93 #include <sys/proc.h> 94 #include <sys/socketvar.h> 95 #include <sys/signalvar.h> 96 #include <sys/uio.h> 97 #include <sys/kernel.h> 98 #include <sys/lwp.h> 99 #include <sys/poll.h> 100 #include <sys/mount.h> 101 #include <sys/syscallargs.h> 102 #include <sys/cpu.h> 103 #include <sys/atomic.h> 104 #include <sys/socketvar.h> 105 #include <sys/sleepq.h> 106 #include <sys/sysctl.h> 107 #include <sys/bitops.h> 108 109 /* Flags for lwp::l_selflag. */ 110 #define SEL_RESET 0 /* awoken, interrupted, or not yet polling */ 111 #define SEL_SCANNING 1 /* polling descriptors */ 112 #define SEL_BLOCKING 2 /* blocking and waiting for event */ 113 #define SEL_EVENT 3 /* interrupted, events set directly */ 114 115 /* 116 * Per-cluster state for select()/poll(). For a system with fewer 117 * than 64 CPUs, this gives us per-CPU clusters. 118 */ 119 #define SELCLUSTERS 64 120 #define SELCLUSTERMASK (SELCLUSTERS - 1) 121 122 typedef struct selcluster { 123 kmutex_t *sc_lock; 124 sleepq_t sc_sleepq; 125 uint64_t sc_mask; 126 int sc_ncoll; 127 } selcluster_t; 128 129 static inline int selscan(char *, const int, const size_t, register_t *); 130 static inline int pollscan(struct pollfd *, const int, register_t *); 131 static void selclear(void); 132 133 static const int sel_flag[] = { 134 POLLRDNORM | POLLHUP | POLLERR, 135 POLLWRNORM | POLLHUP | POLLERR, 136 POLLRDBAND 137 }; 138 139 /* 140 * LWPs are woken using the sleep queue only due to a collision, the case 141 * with the maximum Suck Factor. Save the cost of sorting for named waiters 142 * by inserting in LIFO order. In the future it would be preferable to not 143 * enqueue LWPs at all, unless subject to a collision. 144 */ 145 syncobj_t select_sobj = { 146 .sobj_flag = SOBJ_SLEEPQ_LIFO, 147 .sobj_unsleep = sleepq_unsleep, 148 .sobj_changepri = sleepq_changepri, 149 .sobj_lendpri = sleepq_lendpri, 150 .sobj_owner = syncobj_noowner, 151 }; 152 153 static selcluster_t *selcluster[SELCLUSTERS] __read_mostly; 154 static int direct_select __read_mostly = 0; 155 156 /* Operations: either select() or poll(). */ 157 const char selop_select[] = "select"; 158 const char selop_poll[] = "poll"; 159 160 /* 161 * Select system call. 162 */ 163 int 164 sys___pselect50(struct lwp *l, const struct sys___pselect50_args *uap, 165 register_t *retval) 166 { 167 /* { 168 syscallarg(int) nd; 169 syscallarg(fd_set *) in; 170 syscallarg(fd_set *) ou; 171 syscallarg(fd_set *) ex; 172 syscallarg(const struct timespec *) ts; 173 syscallarg(sigset_t *) mask; 174 } */ 175 struct timespec ats, *ts = NULL; 176 sigset_t amask, *mask = NULL; 177 int error; 178 179 if (SCARG(uap, ts)) { 180 error = copyin(SCARG(uap, ts), &ats, sizeof(ats)); 181 if (error) 182 return error; 183 ts = &ats; 184 } 185 if (SCARG(uap, mask) != NULL) { 186 error = copyin(SCARG(uap, mask), &amask, sizeof(amask)); 187 if (error) 188 return error; 189 mask = &amask; 190 } 191 192 return selcommon(retval, SCARG(uap, nd), SCARG(uap, in), 193 SCARG(uap, ou), SCARG(uap, ex), ts, mask); 194 } 195 196 int 197 sys___select50(struct lwp *l, const struct sys___select50_args *uap, 198 register_t *retval) 199 { 200 /* { 201 syscallarg(int) nd; 202 syscallarg(fd_set *) in; 203 syscallarg(fd_set *) ou; 204 syscallarg(fd_set *) ex; 205 syscallarg(struct timeval *) tv; 206 } */ 207 struct timeval atv; 208 struct timespec ats, *ts = NULL; 209 int error; 210 211 if (SCARG(uap, tv)) { 212 error = copyin(SCARG(uap, tv), (void *)&atv, sizeof(atv)); 213 if (error) 214 return error; 215 216 if (atv.tv_usec < 0 || atv.tv_usec >= 1000000) 217 return EINVAL; 218 219 TIMEVAL_TO_TIMESPEC(&atv, &ats); 220 ts = &ats; 221 } 222 223 return selcommon(retval, SCARG(uap, nd), SCARG(uap, in), 224 SCARG(uap, ou), SCARG(uap, ex), ts, NULL); 225 } 226 227 /* 228 * sel_do_scan: common code to perform the scan on descriptors. 229 */ 230 static int 231 sel_do_scan(const char *opname, void *fds, const int nf, const size_t ni, 232 struct timespec *ts, sigset_t *mask, register_t *retval) 233 { 234 lwp_t * const l = curlwp; 235 selcluster_t *sc; 236 kmutex_t *lock; 237 struct timespec sleepts; 238 int error, timo; 239 240 timo = 0; 241 if (ts && inittimeleft(ts, &sleepts) == -1) { 242 return EINVAL; 243 } 244 245 if (__predict_false(mask)) 246 sigsuspendsetup(l, mask); 247 248 /* 249 * We may context switch during or at any time after picking a CPU 250 * and cluster to associate with, but it doesn't matter. In the 251 * unlikely event we migrate elsewhere all we risk is a little lock 252 * contention; correctness is not sacrificed. 253 */ 254 sc = curcpu()->ci_data.cpu_selcluster; 255 lock = sc->sc_lock; 256 l->l_selcluster = sc; 257 258 if (opname == selop_select) { 259 l->l_selbits = fds; 260 l->l_selni = ni; 261 } else { 262 l->l_selbits = NULL; 263 } 264 265 for (;;) { 266 int ncoll; 267 268 SLIST_INIT(&l->l_selwait); 269 l->l_selret = 0; 270 271 /* 272 * No need to lock. If this is overwritten by another value 273 * while scanning, we will retry below. We only need to see 274 * exact state from the descriptors that we are about to poll, 275 * and lock activity resulting from fo_poll is enough to 276 * provide an up to date value for new polling activity. 277 */ 278 if (ts && (ts->tv_sec | ts->tv_nsec | direct_select) == 0) { 279 /* Non-blocking: no need for selrecord()/selclear() */ 280 l->l_selflag = SEL_RESET; 281 } else { 282 l->l_selflag = SEL_SCANNING; 283 } 284 ncoll = sc->sc_ncoll; 285 membar_exit(); 286 287 if (opname == selop_select) { 288 error = selscan((char *)fds, nf, ni, retval); 289 } else { 290 error = pollscan((struct pollfd *)fds, nf, retval); 291 } 292 if (error || *retval) 293 break; 294 if (ts && (timo = gettimeleft(ts, &sleepts)) <= 0) 295 break; 296 /* 297 * Acquire the lock and perform the (re)checks. Note, if 298 * collision has occurred, then our state does not matter, 299 * as we must perform re-scan. Therefore, check it first. 300 */ 301 state_check: 302 mutex_spin_enter(lock); 303 if (__predict_false(sc->sc_ncoll != ncoll)) { 304 /* Collision: perform re-scan. */ 305 mutex_spin_exit(lock); 306 selclear(); 307 continue; 308 } 309 if (__predict_true(l->l_selflag == SEL_EVENT)) { 310 /* Events occurred, they are set directly. */ 311 mutex_spin_exit(lock); 312 break; 313 } 314 if (__predict_true(l->l_selflag == SEL_RESET)) { 315 /* Events occurred, but re-scan is requested. */ 316 mutex_spin_exit(lock); 317 selclear(); 318 continue; 319 } 320 /* Nothing happen, therefore - sleep. */ 321 l->l_selflag = SEL_BLOCKING; 322 l->l_kpriority = true; 323 sleepq_enter(&sc->sc_sleepq, l, lock); 324 sleepq_enqueue(&sc->sc_sleepq, sc, opname, &select_sobj, true); 325 error = sleepq_block(timo, true); 326 if (error != 0) { 327 break; 328 } 329 /* Awoken: need to check the state. */ 330 goto state_check; 331 } 332 selclear(); 333 334 /* Add direct events if any. */ 335 if (l->l_selflag == SEL_EVENT) { 336 KASSERT(l->l_selret != 0); 337 *retval += l->l_selret; 338 } 339 340 if (__predict_false(mask)) 341 sigsuspendteardown(l); 342 343 /* select and poll are not restarted after signals... */ 344 if (error == ERESTART) 345 return EINTR; 346 if (error == EWOULDBLOCK) 347 return 0; 348 return error; 349 } 350 351 int 352 selcommon(register_t *retval, int nd, fd_set *u_in, fd_set *u_ou, 353 fd_set *u_ex, struct timespec *ts, sigset_t *mask) 354 { 355 char smallbits[howmany(FD_SETSIZE, NFDBITS) * 356 sizeof(fd_mask) * 6]; 357 char *bits; 358 int error, nf; 359 size_t ni; 360 361 if (nd < 0) 362 return (EINVAL); 363 nf = atomic_load_consume(&curlwp->l_fd->fd_dt)->dt_nfiles; 364 if (nd > nf) { 365 /* forgiving; slightly wrong */ 366 nd = nf; 367 } 368 ni = howmany(nd, NFDBITS) * sizeof(fd_mask); 369 if (ni * 6 > sizeof(smallbits)) 370 bits = kmem_alloc(ni * 6, KM_SLEEP); 371 else 372 bits = smallbits; 373 374 #define getbits(name, x) \ 375 if (u_ ## name) { \ 376 error = copyin(u_ ## name, bits + ni * x, ni); \ 377 if (error) \ 378 goto fail; \ 379 } else \ 380 memset(bits + ni * x, 0, ni); 381 getbits(in, 0); 382 getbits(ou, 1); 383 getbits(ex, 2); 384 #undef getbits 385 386 error = sel_do_scan(selop_select, bits, nd, ni, ts, mask, retval); 387 if (error == 0 && u_in != NULL) 388 error = copyout(bits + ni * 3, u_in, ni); 389 if (error == 0 && u_ou != NULL) 390 error = copyout(bits + ni * 4, u_ou, ni); 391 if (error == 0 && u_ex != NULL) 392 error = copyout(bits + ni * 5, u_ex, ni); 393 fail: 394 if (bits != smallbits) 395 kmem_free(bits, ni * 6); 396 return (error); 397 } 398 399 static inline int 400 selscan(char *bits, const int nfd, const size_t ni, register_t *retval) 401 { 402 fd_mask *ibitp, *obitp; 403 int msk, i, j, fd, n; 404 file_t *fp; 405 lwp_t *l; 406 407 ibitp = (fd_mask *)(bits + ni * 0); 408 obitp = (fd_mask *)(bits + ni * 3); 409 n = 0; 410 l = curlwp; 411 412 memset(obitp, 0, ni * 3); 413 for (msk = 0; msk < 3; msk++) { 414 for (i = 0; i < nfd; i += NFDBITS) { 415 fd_mask ibits, obits; 416 417 ibits = *ibitp; 418 obits = 0; 419 while ((j = ffs(ibits)) && (fd = i + --j) < nfd) { 420 ibits &= ~(1U << j); 421 if ((fp = fd_getfile(fd)) == NULL) 422 return (EBADF); 423 /* 424 * Setup an argument to selrecord(), which is 425 * a file descriptor number. 426 */ 427 l->l_selrec = fd; 428 if ((*fp->f_ops->fo_poll)(fp, sel_flag[msk])) { 429 if (!direct_select) { 430 /* 431 * Have events: do nothing in 432 * selrecord(). 433 */ 434 l->l_selflag = SEL_RESET; 435 } 436 obits |= (1U << j); 437 n++; 438 } 439 fd_putfile(fd); 440 } 441 if (obits != 0) { 442 if (direct_select) { 443 kmutex_t *lock; 444 lock = l->l_selcluster->sc_lock; 445 mutex_spin_enter(lock); 446 *obitp |= obits; 447 mutex_spin_exit(lock); 448 } else { 449 *obitp |= obits; 450 } 451 } 452 ibitp++; 453 obitp++; 454 } 455 } 456 *retval = n; 457 return (0); 458 } 459 460 /* 461 * Poll system call. 462 */ 463 int 464 sys_poll(struct lwp *l, const struct sys_poll_args *uap, register_t *retval) 465 { 466 /* { 467 syscallarg(struct pollfd *) fds; 468 syscallarg(u_int) nfds; 469 syscallarg(int) timeout; 470 } */ 471 struct timespec ats, *ts = NULL; 472 473 if (SCARG(uap, timeout) != INFTIM) { 474 ats.tv_sec = SCARG(uap, timeout) / 1000; 475 ats.tv_nsec = (SCARG(uap, timeout) % 1000) * 1000000; 476 ts = &ats; 477 } 478 479 return pollcommon(retval, SCARG(uap, fds), SCARG(uap, nfds), ts, NULL); 480 } 481 482 /* 483 * Poll system call. 484 */ 485 int 486 sys___pollts50(struct lwp *l, const struct sys___pollts50_args *uap, 487 register_t *retval) 488 { 489 /* { 490 syscallarg(struct pollfd *) fds; 491 syscallarg(u_int) nfds; 492 syscallarg(const struct timespec *) ts; 493 syscallarg(const sigset_t *) mask; 494 } */ 495 struct timespec ats, *ts = NULL; 496 sigset_t amask, *mask = NULL; 497 int error; 498 499 if (SCARG(uap, ts)) { 500 error = copyin(SCARG(uap, ts), &ats, sizeof(ats)); 501 if (error) 502 return error; 503 ts = &ats; 504 } 505 if (SCARG(uap, mask)) { 506 error = copyin(SCARG(uap, mask), &amask, sizeof(amask)); 507 if (error) 508 return error; 509 mask = &amask; 510 } 511 512 return pollcommon(retval, SCARG(uap, fds), SCARG(uap, nfds), ts, mask); 513 } 514 515 int 516 pollcommon(register_t *retval, struct pollfd *u_fds, u_int nfds, 517 struct timespec *ts, sigset_t *mask) 518 { 519 struct pollfd smallfds[32]; 520 struct pollfd *fds; 521 int error; 522 size_t ni; 523 524 if (nfds > curlwp->l_proc->p_rlimit[RLIMIT_NOFILE].rlim_max + 1000) { 525 /* 526 * Prevent userland from causing over-allocation. 527 * Raising the default limit too high can still cause 528 * a lot of memory to be allocated, but this also means 529 * that the file descriptor array will also be large. 530 * 531 * To reduce the memory requirements here, we could 532 * process the 'fds' array in chunks, but that 533 * is a lot of code that isn't normally useful. 534 * (Or just move the copyin/out into pollscan().) 535 * 536 * Historically the code silently truncated 'fds' to 537 * dt_nfiles entries - but that does cause issues. 538 * 539 * Using the max limit equivalent to sysctl 540 * kern.maxfiles is the moral equivalent of OPEN_MAX 541 * as specified by POSIX. 542 * 543 * We add a slop of 1000 in case the resource limit was 544 * changed after opening descriptors or the same descriptor 545 * was specified more than once. 546 */ 547 return EINVAL; 548 } 549 ni = nfds * sizeof(struct pollfd); 550 if (ni > sizeof(smallfds)) 551 fds = kmem_alloc(ni, KM_SLEEP); 552 else 553 fds = smallfds; 554 555 error = copyin(u_fds, fds, ni); 556 if (error) 557 goto fail; 558 559 error = sel_do_scan(selop_poll, fds, nfds, ni, ts, mask, retval); 560 if (error == 0) 561 error = copyout(fds, u_fds, ni); 562 fail: 563 if (fds != smallfds) 564 kmem_free(fds, ni); 565 return (error); 566 } 567 568 static inline int 569 pollscan(struct pollfd *fds, const int nfd, register_t *retval) 570 { 571 file_t *fp; 572 int i, n = 0, revents; 573 574 for (i = 0; i < nfd; i++, fds++) { 575 fds->revents = 0; 576 if (fds->fd < 0) { 577 revents = 0; 578 } else if ((fp = fd_getfile(fds->fd)) == NULL) { 579 revents = POLLNVAL; 580 } else { 581 /* 582 * Perform poll: registers select request or returns 583 * the events which are set. Setup an argument for 584 * selrecord(), which is a pointer to struct pollfd. 585 */ 586 curlwp->l_selrec = (uintptr_t)fds; 587 revents = (*fp->f_ops->fo_poll)(fp, 588 fds->events | POLLERR | POLLHUP); 589 fd_putfile(fds->fd); 590 } 591 if (revents) { 592 if (!direct_select) { 593 /* Have events: do nothing in selrecord(). */ 594 curlwp->l_selflag = SEL_RESET; 595 } 596 fds->revents = revents; 597 n++; 598 } 599 } 600 *retval = n; 601 return (0); 602 } 603 604 int 605 seltrue(dev_t dev, int events, lwp_t *l) 606 { 607 608 return (events & (POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM)); 609 } 610 611 /* 612 * Record a select request. Concurrency issues: 613 * 614 * The caller holds the same lock across calls to selrecord() and 615 * selnotify(), so we don't need to consider a concurrent wakeup 616 * while in this routine. 617 * 618 * The only activity we need to guard against is selclear(), called by 619 * another thread that is exiting sel_do_scan(). 620 * `sel_lwp' can only become non-NULL while the caller's lock is held, 621 * so it cannot become non-NULL due to a change made by another thread 622 * while we are in this routine. It can only become _NULL_ due to a 623 * call to selclear(). 624 * 625 * If it is non-NULL and != selector there is the potential for 626 * selclear() to be called by another thread. If either of those 627 * conditions are true, we're not interested in touching the `named 628 * waiter' part of the selinfo record because we need to record a 629 * collision. Hence there is no need for additional locking in this 630 * routine. 631 */ 632 void 633 selrecord(lwp_t *selector, struct selinfo *sip) 634 { 635 selcluster_t *sc; 636 lwp_t *other; 637 638 KASSERT(selector == curlwp); 639 640 sc = selector->l_selcluster; 641 other = sip->sel_lwp; 642 643 if (selector->l_selflag == SEL_RESET) { 644 /* 0. We're not going to block - will poll again if needed. */ 645 } else if (other == selector) { 646 /* 1. We (selector) already claimed to be the first LWP. */ 647 KASSERT(sip->sel_cluster == sc); 648 } else if (other == NULL) { 649 /* 650 * 2. No first LWP, therefore we (selector) are the first. 651 * 652 * There may be unnamed waiters (collisions). Issue a memory 653 * barrier to ensure that we access sel_lwp (above) before 654 * other fields - this guards against a call to selclear(). 655 */ 656 membar_enter(); 657 sip->sel_lwp = selector; 658 SLIST_INSERT_HEAD(&selector->l_selwait, sip, sel_chain); 659 /* Copy the argument, which is for selnotify(). */ 660 sip->sel_fdinfo = selector->l_selrec; 661 /* Replace selinfo's lock with the chosen cluster's lock. */ 662 sip->sel_cluster = sc; 663 } else { 664 /* 3. Multiple waiters: record a collision. */ 665 sip->sel_collision |= sc->sc_mask; 666 KASSERT(sip->sel_cluster != NULL); 667 } 668 } 669 670 /* 671 * Record a knote. 672 * 673 * The caller holds the same lock as for selrecord(). 674 */ 675 void 676 selrecord_knote(struct selinfo *sip, struct knote *kn) 677 { 678 SLIST_INSERT_HEAD(&sip->sel_klist, kn, kn_selnext); 679 } 680 681 /* 682 * Remove a knote. 683 * 684 * The caller holds the same lock as for selrecord(). 685 * 686 * Returns true if the last knote was removed and the list 687 * is now empty. 688 */ 689 bool 690 selremove_knote(struct selinfo *sip, struct knote *kn) 691 { 692 SLIST_REMOVE(&sip->sel_klist, kn, knote, kn_selnext); 693 return SLIST_EMPTY(&sip->sel_klist); 694 } 695 696 /* 697 * sel_setevents: a helper function for selnotify(), to set the events 698 * for LWP sleeping in selcommon() or pollcommon(). 699 */ 700 static inline bool 701 sel_setevents(lwp_t *l, struct selinfo *sip, const int events) 702 { 703 const int oflag = l->l_selflag; 704 int ret = 0; 705 706 /* 707 * If we require re-scan or it was required by somebody else, 708 * then just (re)set SEL_RESET and return. 709 */ 710 if (__predict_false(events == 0 || oflag == SEL_RESET)) { 711 l->l_selflag = SEL_RESET; 712 return true; 713 } 714 /* 715 * Direct set. Note: select state of LWP is locked. First, 716 * determine whether it is selcommon() or pollcommon(). 717 */ 718 if (l->l_selbits != NULL) { 719 const size_t ni = l->l_selni; 720 fd_mask *fds = (fd_mask *)l->l_selbits; 721 fd_mask *ofds = (fd_mask *)((char *)fds + ni * 3); 722 const int fd = sip->sel_fdinfo, fbit = 1 << (fd & __NFDMASK); 723 const int idx = fd >> __NFDSHIFT; 724 int n; 725 726 for (n = 0; n < 3; n++) { 727 if ((fds[idx] & fbit) != 0 && 728 (ofds[idx] & fbit) == 0 && 729 (sel_flag[n] & events)) { 730 ofds[idx] |= fbit; 731 ret++; 732 } 733 fds = (fd_mask *)((char *)fds + ni); 734 ofds = (fd_mask *)((char *)ofds + ni); 735 } 736 } else { 737 struct pollfd *pfd = (void *)sip->sel_fdinfo; 738 int revents = events & (pfd->events | POLLERR | POLLHUP); 739 740 if (revents) { 741 if (pfd->revents == 0) 742 ret = 1; 743 pfd->revents |= revents; 744 } 745 } 746 /* Check whether there are any events to return. */ 747 if (!ret) { 748 return false; 749 } 750 /* Indicate direct set and note the event (cluster lock is held). */ 751 l->l_selflag = SEL_EVENT; 752 l->l_selret += ret; 753 return true; 754 } 755 756 /* 757 * Do a wakeup when a selectable event occurs. Concurrency issues: 758 * 759 * As per selrecord(), the caller's object lock is held. If there 760 * is a named waiter, we must acquire the associated selcluster's lock 761 * in order to synchronize with selclear() and pollers going to sleep 762 * in sel_do_scan(). 763 * 764 * sip->sel_cluser cannot change at this point, as it is only changed 765 * in selrecord(), and concurrent calls to selrecord() are locked 766 * out by the caller. 767 */ 768 void 769 selnotify(struct selinfo *sip, int events, long knhint) 770 { 771 selcluster_t *sc; 772 uint64_t mask; 773 int index, oflag; 774 lwp_t *l; 775 kmutex_t *lock; 776 777 KNOTE(&sip->sel_klist, knhint); 778 779 if (sip->sel_lwp != NULL) { 780 /* One named LWP is waiting. */ 781 sc = sip->sel_cluster; 782 lock = sc->sc_lock; 783 mutex_spin_enter(lock); 784 /* Still there? */ 785 if (sip->sel_lwp != NULL) { 786 /* 787 * Set the events for our LWP and indicate that. 788 * Otherwise, request for a full re-scan. 789 */ 790 l = sip->sel_lwp; 791 oflag = l->l_selflag; 792 793 if (!direct_select) { 794 l->l_selflag = SEL_RESET; 795 } else if (!sel_setevents(l, sip, events)) { 796 /* No events to return. */ 797 mutex_spin_exit(lock); 798 return; 799 } 800 801 /* 802 * If thread is sleeping, wake it up. If it's not 803 * yet asleep, it will notice the change in state 804 * and will re-poll the descriptors. 805 */ 806 if (oflag == SEL_BLOCKING && l->l_mutex == lock) { 807 KASSERT(l->l_wchan == sc); 808 sleepq_unsleep(l, false); 809 } 810 } 811 mutex_spin_exit(lock); 812 } 813 814 if ((mask = sip->sel_collision) != 0) { 815 /* 816 * There was a collision (multiple waiters): we must 817 * inform all potentially interested waiters. 818 */ 819 sip->sel_collision = 0; 820 do { 821 index = ffs64(mask) - 1; 822 mask ^= __BIT(index); 823 sc = selcluster[index]; 824 lock = sc->sc_lock; 825 mutex_spin_enter(lock); 826 sc->sc_ncoll++; 827 sleepq_wake(&sc->sc_sleepq, sc, (u_int)-1, lock); 828 } while (__predict_false(mask != 0)); 829 } 830 } 831 832 /* 833 * Remove an LWP from all objects that it is waiting for. Concurrency 834 * issues: 835 * 836 * The object owner's (e.g. device driver) lock is not held here. Calls 837 * can be made to selrecord() and we do not synchronize against those 838 * directly using locks. However, we use `sel_lwp' to lock out changes. 839 * Before clearing it we must use memory barriers to ensure that we can 840 * safely traverse the list of selinfo records. 841 */ 842 static void 843 selclear(void) 844 { 845 struct selinfo *sip, *next; 846 selcluster_t *sc; 847 lwp_t *l; 848 kmutex_t *lock; 849 850 l = curlwp; 851 sc = l->l_selcluster; 852 lock = sc->sc_lock; 853 854 /* 855 * If the request was non-blocking, or we found events on the first 856 * descriptor, there will be no need to clear anything - avoid 857 * taking the lock. 858 */ 859 if (SLIST_EMPTY(&l->l_selwait)) { 860 return; 861 } 862 863 mutex_spin_enter(lock); 864 for (sip = SLIST_FIRST(&l->l_selwait); sip != NULL; sip = next) { 865 KASSERT(sip->sel_lwp == l); 866 KASSERT(sip->sel_cluster == l->l_selcluster); 867 868 /* 869 * Read link to next selinfo record, if any. 870 * It's no longer safe to touch `sip' after clearing 871 * `sel_lwp', so ensure that the read of `sel_chain' 872 * completes before the clearing of sel_lwp becomes 873 * globally visible. 874 */ 875 next = SLIST_NEXT(sip, sel_chain); 876 membar_exit(); 877 /* Release the record for another named waiter to use. */ 878 sip->sel_lwp = NULL; 879 } 880 mutex_spin_exit(lock); 881 } 882 883 /* 884 * Initialize the select/poll system calls. Called once for each 885 * CPU in the system, as they are attached. 886 */ 887 void 888 selsysinit(struct cpu_info *ci) 889 { 890 selcluster_t *sc; 891 u_int index; 892 893 /* If already a cluster in place for this bit, re-use. */ 894 index = cpu_index(ci) & SELCLUSTERMASK; 895 sc = selcluster[index]; 896 if (sc == NULL) { 897 sc = kmem_alloc(roundup2(sizeof(selcluster_t), 898 coherency_unit) + coherency_unit, KM_SLEEP); 899 sc = (void *)roundup2((uintptr_t)sc, coherency_unit); 900 sc->sc_lock = mutex_obj_alloc(MUTEX_DEFAULT, IPL_SCHED); 901 sleepq_init(&sc->sc_sleepq); 902 sc->sc_ncoll = 0; 903 sc->sc_mask = __BIT(index); 904 selcluster[index] = sc; 905 } 906 ci->ci_data.cpu_selcluster = sc; 907 } 908 909 /* 910 * Initialize a selinfo record. 911 */ 912 void 913 selinit(struct selinfo *sip) 914 { 915 916 memset(sip, 0, sizeof(*sip)); 917 } 918 919 /* 920 * Destroy a selinfo record. The owning object must not gain new 921 * references while this is in progress: all activity on the record 922 * must be stopped. 923 * 924 * Concurrency issues: we only need guard against a call to selclear() 925 * by a thread exiting sel_do_scan(). The caller has prevented further 926 * references being made to the selinfo record via selrecord(), and it 927 * will not call selnotify() again. 928 */ 929 void 930 seldestroy(struct selinfo *sip) 931 { 932 selcluster_t *sc; 933 kmutex_t *lock; 934 lwp_t *l; 935 936 if (sip->sel_lwp == NULL) 937 return; 938 939 /* 940 * Lock out selclear(). The selcluster pointer can't change while 941 * we are here since it is only ever changed in selrecord(), 942 * and that will not be entered again for this record because 943 * it is dying. 944 */ 945 KASSERT(sip->sel_cluster != NULL); 946 sc = sip->sel_cluster; 947 lock = sc->sc_lock; 948 mutex_spin_enter(lock); 949 if ((l = sip->sel_lwp) != NULL) { 950 /* 951 * This should rarely happen, so although SLIST_REMOVE() 952 * is slow, using it here is not a problem. 953 */ 954 KASSERT(l->l_selcluster == sc); 955 SLIST_REMOVE(&l->l_selwait, sip, selinfo, sel_chain); 956 sip->sel_lwp = NULL; 957 } 958 mutex_spin_exit(lock); 959 } 960 961 /* 962 * System control nodes. 963 */ 964 SYSCTL_SETUP(sysctl_select_setup, "sysctl select setup") 965 { 966 967 sysctl_createv(clog, 0, NULL, NULL, 968 CTLFLAG_PERMANENT | CTLFLAG_READWRITE, 969 CTLTYPE_INT, "direct_select", 970 SYSCTL_DESCR("Enable/disable direct select (for testing)"), 971 NULL, 0, &direct_select, 0, 972 CTL_KERN, CTL_CREATE, CTL_EOL); 973 } 974