1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * This file contains the procedures for the handling of select and poll 4 * 5 * Created for Linux based loosely upon Mathius Lattner's minix 6 * patches by Peter MacDonald. Heavily edited by Linus. 7 * 8 * 4 February 1994 9 * COFF/ELF binary emulation. If the process has the STICKY_TIMEOUTS 10 * flag set in its personality we do *not* modify the given timeout 11 * parameter to reflect time remaining. 12 * 13 * 24 January 2000 14 * Changed sys_poll()/do_poll() to use PAGE_SIZE chunk-based allocation 15 * of fds to overcome nfds < 16390 descriptors limit (Tigran Aivazian). 16 */ 17 18 #include <linux/kernel.h> 19 #include <linux/sched/signal.h> 20 #include <linux/sched/rt.h> 21 #include <linux/syscalls.h> 22 #include <linux/export.h> 23 #include <linux/slab.h> 24 #include <linux/poll.h> 25 #include <linux/personality.h> /* for STICKY_TIMEOUTS */ 26 #include <linux/file.h> 27 #include <linux/fdtable.h> 28 #include <linux/fs.h> 29 #include <linux/rcupdate.h> 30 #include <linux/hrtimer.h> 31 #include <linux/freezer.h> 32 #include <net/busy_poll.h> 33 #include <linux/vmalloc.h> 34 35 #include <linux/uaccess.h> 36 37 38 /* 39 * Estimate expected accuracy in ns from a timeval. 40 * 41 * After quite a bit of churning around, we've settled on 42 * a simple thing of taking 0.1% of the timeout as the 43 * slack, with a cap of 100 msec. 44 * "nice" tasks get a 0.5% slack instead. 45 * 46 * Consider this comment an open invitation to come up with even 47 * better solutions.. 48 */ 49 50 #define MAX_SLACK (100 * NSEC_PER_MSEC) 51 52 static long __estimate_accuracy(struct timespec64 *tv) 53 { 54 long slack; 55 int divfactor = 1000; 56 57 if (tv->tv_sec < 0) 58 return 0; 59 60 if (task_nice(current) > 0) 61 divfactor = divfactor / 5; 62 63 if (tv->tv_sec > MAX_SLACK / (NSEC_PER_SEC/divfactor)) 64 return MAX_SLACK; 65 66 slack = tv->tv_nsec / divfactor; 67 slack += tv->tv_sec * (NSEC_PER_SEC/divfactor); 68 69 if (slack > MAX_SLACK) 70 return MAX_SLACK; 71 72 return slack; 73 } 74 75 u64 select_estimate_accuracy(struct timespec64 *tv) 76 { 77 u64 ret; 78 struct timespec64 now; 79 80 /* 81 * Realtime tasks get a slack of 0 for obvious reasons. 82 */ 83 84 if (rt_task(current)) 85 return 0; 86 87 ktime_get_ts64(&now); 88 now = timespec64_sub(*tv, now); 89 ret = __estimate_accuracy(&now); 90 if (ret < current->timer_slack_ns) 91 return current->timer_slack_ns; 92 return ret; 93 } 94 95 96 97 struct poll_table_page { 98 struct poll_table_page * next; 99 struct poll_table_entry * entry; 100 struct poll_table_entry entries[0]; 101 }; 102 103 #define POLL_TABLE_FULL(table) \ 104 ((unsigned long)((table)->entry+1) > PAGE_SIZE + (unsigned long)(table)) 105 106 /* 107 * Ok, Peter made a complicated, but straightforward multiple_wait() function. 108 * I have rewritten this, taking some shortcuts: This code may not be easy to 109 * follow, but it should be free of race-conditions, and it's practical. If you 110 * understand what I'm doing here, then you understand how the linux 111 * sleep/wakeup mechanism works. 112 * 113 * Two very simple procedures, poll_wait() and poll_freewait() make all the 114 * work. poll_wait() is an inline-function defined in <linux/poll.h>, 115 * as all select/poll functions have to call it to add an entry to the 116 * poll table. 117 */ 118 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address, 119 poll_table *p); 120 121 void poll_initwait(struct poll_wqueues *pwq) 122 { 123 init_poll_funcptr(&pwq->pt, __pollwait); 124 pwq->polling_task = current; 125 pwq->triggered = 0; 126 pwq->error = 0; 127 pwq->table = NULL; 128 pwq->inline_index = 0; 129 } 130 EXPORT_SYMBOL(poll_initwait); 131 132 static void free_poll_entry(struct poll_table_entry *entry) 133 { 134 remove_wait_queue(entry->wait_address, &entry->wait); 135 fput(entry->filp); 136 } 137 138 void poll_freewait(struct poll_wqueues *pwq) 139 { 140 struct poll_table_page * p = pwq->table; 141 int i; 142 for (i = 0; i < pwq->inline_index; i++) 143 free_poll_entry(pwq->inline_entries + i); 144 while (p) { 145 struct poll_table_entry * entry; 146 struct poll_table_page *old; 147 148 entry = p->entry; 149 do { 150 entry--; 151 free_poll_entry(entry); 152 } while (entry > p->entries); 153 old = p; 154 p = p->next; 155 free_page((unsigned long) old); 156 } 157 } 158 EXPORT_SYMBOL(poll_freewait); 159 160 static struct poll_table_entry *poll_get_entry(struct poll_wqueues *p) 161 { 162 struct poll_table_page *table = p->table; 163 164 if (p->inline_index < N_INLINE_POLL_ENTRIES) 165 return p->inline_entries + p->inline_index++; 166 167 if (!table || POLL_TABLE_FULL(table)) { 168 struct poll_table_page *new_table; 169 170 new_table = (struct poll_table_page *) __get_free_page(GFP_KERNEL); 171 if (!new_table) { 172 p->error = -ENOMEM; 173 return NULL; 174 } 175 new_table->entry = new_table->entries; 176 new_table->next = table; 177 p->table = new_table; 178 table = new_table; 179 } 180 181 return table->entry++; 182 } 183 184 static int __pollwake(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) 185 { 186 struct poll_wqueues *pwq = wait->private; 187 DECLARE_WAITQUEUE(dummy_wait, pwq->polling_task); 188 189 /* 190 * Although this function is called under waitqueue lock, LOCK 191 * doesn't imply write barrier and the users expect write 192 * barrier semantics on wakeup functions. The following 193 * smp_wmb() is equivalent to smp_wmb() in try_to_wake_up() 194 * and is paired with smp_store_mb() in poll_schedule_timeout. 195 */ 196 smp_wmb(); 197 pwq->triggered = 1; 198 199 /* 200 * Perform the default wake up operation using a dummy 201 * waitqueue. 202 * 203 * TODO: This is hacky but there currently is no interface to 204 * pass in @sync. @sync is scheduled to be removed and once 205 * that happens, wake_up_process() can be used directly. 206 */ 207 return default_wake_function(&dummy_wait, mode, sync, key); 208 } 209 210 static int pollwake(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) 211 { 212 struct poll_table_entry *entry; 213 214 entry = container_of(wait, struct poll_table_entry, wait); 215 if (key && !(key_to_poll(key) & entry->key)) 216 return 0; 217 return __pollwake(wait, mode, sync, key); 218 } 219 220 /* Add a new entry */ 221 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address, 222 poll_table *p) 223 { 224 struct poll_wqueues *pwq = container_of(p, struct poll_wqueues, pt); 225 struct poll_table_entry *entry = poll_get_entry(pwq); 226 if (!entry) 227 return; 228 entry->filp = get_file(filp); 229 entry->wait_address = wait_address; 230 entry->key = p->_key; 231 init_waitqueue_func_entry(&entry->wait, pollwake); 232 entry->wait.private = pwq; 233 add_wait_queue(wait_address, &entry->wait); 234 } 235 236 static int poll_schedule_timeout(struct poll_wqueues *pwq, int state, 237 ktime_t *expires, unsigned long slack) 238 { 239 int rc = -EINTR; 240 241 set_current_state(state); 242 if (!pwq->triggered) 243 rc = schedule_hrtimeout_range(expires, slack, HRTIMER_MODE_ABS); 244 __set_current_state(TASK_RUNNING); 245 246 /* 247 * Prepare for the next iteration. 248 * 249 * The following smp_store_mb() serves two purposes. First, it's 250 * the counterpart rmb of the wmb in pollwake() such that data 251 * written before wake up is always visible after wake up. 252 * Second, the full barrier guarantees that triggered clearing 253 * doesn't pass event check of the next iteration. Note that 254 * this problem doesn't exist for the first iteration as 255 * add_wait_queue() has full barrier semantics. 256 */ 257 smp_store_mb(pwq->triggered, 0); 258 259 return rc; 260 } 261 262 /** 263 * poll_select_set_timeout - helper function to setup the timeout value 264 * @to: pointer to timespec64 variable for the final timeout 265 * @sec: seconds (from user space) 266 * @nsec: nanoseconds (from user space) 267 * 268 * Note, we do not use a timespec for the user space value here, That 269 * way we can use the function for timeval and compat interfaces as well. 270 * 271 * Returns -EINVAL if sec/nsec are not normalized. Otherwise 0. 272 */ 273 int poll_select_set_timeout(struct timespec64 *to, time64_t sec, long nsec) 274 { 275 struct timespec64 ts = {.tv_sec = sec, .tv_nsec = nsec}; 276 277 if (!timespec64_valid(&ts)) 278 return -EINVAL; 279 280 /* Optimize for the zero timeout value here */ 281 if (!sec && !nsec) { 282 to->tv_sec = to->tv_nsec = 0; 283 } else { 284 ktime_get_ts64(to); 285 *to = timespec64_add_safe(*to, ts); 286 } 287 return 0; 288 } 289 290 static int poll_select_copy_remaining(struct timespec64 *end_time, 291 void __user *p, 292 int timeval, int ret) 293 { 294 struct timespec64 rts; 295 struct timeval rtv; 296 297 if (!p) 298 return ret; 299 300 if (current->personality & STICKY_TIMEOUTS) 301 goto sticky; 302 303 /* No update for zero timeout */ 304 if (!end_time->tv_sec && !end_time->tv_nsec) 305 return ret; 306 307 ktime_get_ts64(&rts); 308 rts = timespec64_sub(*end_time, rts); 309 if (rts.tv_sec < 0) 310 rts.tv_sec = rts.tv_nsec = 0; 311 312 313 if (timeval) { 314 if (sizeof(rtv) > sizeof(rtv.tv_sec) + sizeof(rtv.tv_usec)) 315 memset(&rtv, 0, sizeof(rtv)); 316 rtv.tv_sec = rts.tv_sec; 317 rtv.tv_usec = rts.tv_nsec / NSEC_PER_USEC; 318 319 if (!copy_to_user(p, &rtv, sizeof(rtv))) 320 return ret; 321 322 } else if (!put_timespec64(&rts, p)) 323 return ret; 324 325 /* 326 * If an application puts its timeval in read-only memory, we 327 * don't want the Linux-specific update to the timeval to 328 * cause a fault after the select has completed 329 * successfully. However, because we're not updating the 330 * timeval, we can't restart the system call. 331 */ 332 333 sticky: 334 if (ret == -ERESTARTNOHAND) 335 ret = -EINTR; 336 return ret; 337 } 338 339 /* 340 * Scalable version of the fd_set. 341 */ 342 343 typedef struct { 344 unsigned long *in, *out, *ex; 345 unsigned long *res_in, *res_out, *res_ex; 346 } fd_set_bits; 347 348 /* 349 * How many longwords for "nr" bits? 350 */ 351 #define FDS_BITPERLONG (8*sizeof(long)) 352 #define FDS_LONGS(nr) (((nr)+FDS_BITPERLONG-1)/FDS_BITPERLONG) 353 #define FDS_BYTES(nr) (FDS_LONGS(nr)*sizeof(long)) 354 355 /* 356 * We do a VERIFY_WRITE here even though we are only reading this time: 357 * we'll write to it eventually.. 358 * 359 * Use "unsigned long" accesses to let user-mode fd_set's be long-aligned. 360 */ 361 static inline 362 int get_fd_set(unsigned long nr, void __user *ufdset, unsigned long *fdset) 363 { 364 nr = FDS_BYTES(nr); 365 if (ufdset) 366 return copy_from_user(fdset, ufdset, nr) ? -EFAULT : 0; 367 368 memset(fdset, 0, nr); 369 return 0; 370 } 371 372 static inline unsigned long __must_check 373 set_fd_set(unsigned long nr, void __user *ufdset, unsigned long *fdset) 374 { 375 if (ufdset) 376 return __copy_to_user(ufdset, fdset, FDS_BYTES(nr)); 377 return 0; 378 } 379 380 static inline 381 void zero_fd_set(unsigned long nr, unsigned long *fdset) 382 { 383 memset(fdset, 0, FDS_BYTES(nr)); 384 } 385 386 #define FDS_IN(fds, n) (fds->in + n) 387 #define FDS_OUT(fds, n) (fds->out + n) 388 #define FDS_EX(fds, n) (fds->ex + n) 389 390 #define BITS(fds, n) (*FDS_IN(fds, n)|*FDS_OUT(fds, n)|*FDS_EX(fds, n)) 391 392 static int max_select_fd(unsigned long n, fd_set_bits *fds) 393 { 394 unsigned long *open_fds; 395 unsigned long set; 396 int max; 397 struct fdtable *fdt; 398 399 /* handle last in-complete long-word first */ 400 set = ~(~0UL << (n & (BITS_PER_LONG-1))); 401 n /= BITS_PER_LONG; 402 fdt = files_fdtable(current->files); 403 open_fds = fdt->open_fds + n; 404 max = 0; 405 if (set) { 406 set &= BITS(fds, n); 407 if (set) { 408 if (!(set & ~*open_fds)) 409 goto get_max; 410 return -EBADF; 411 } 412 } 413 while (n) { 414 open_fds--; 415 n--; 416 set = BITS(fds, n); 417 if (!set) 418 continue; 419 if (set & ~*open_fds) 420 return -EBADF; 421 if (max) 422 continue; 423 get_max: 424 do { 425 max++; 426 set >>= 1; 427 } while (set); 428 max += n * BITS_PER_LONG; 429 } 430 431 return max; 432 } 433 434 #define POLLIN_SET (EPOLLRDNORM | EPOLLRDBAND | EPOLLIN | EPOLLHUP | EPOLLERR) 435 #define POLLOUT_SET (EPOLLWRBAND | EPOLLWRNORM | EPOLLOUT | EPOLLERR) 436 #define POLLEX_SET (EPOLLPRI) 437 438 static inline void wait_key_set(poll_table *wait, unsigned long in, 439 unsigned long out, unsigned long bit, 440 __poll_t ll_flag) 441 { 442 wait->_key = POLLEX_SET | ll_flag; 443 if (in & bit) 444 wait->_key |= POLLIN_SET; 445 if (out & bit) 446 wait->_key |= POLLOUT_SET; 447 } 448 449 static int do_select(int n, fd_set_bits *fds, struct timespec64 *end_time) 450 { 451 ktime_t expire, *to = NULL; 452 struct poll_wqueues table; 453 poll_table *wait; 454 int retval, i, timed_out = 0; 455 u64 slack = 0; 456 __poll_t busy_flag = net_busy_loop_on() ? POLL_BUSY_LOOP : 0; 457 unsigned long busy_start = 0; 458 459 rcu_read_lock(); 460 retval = max_select_fd(n, fds); 461 rcu_read_unlock(); 462 463 if (retval < 0) 464 return retval; 465 n = retval; 466 467 poll_initwait(&table); 468 wait = &table.pt; 469 if (end_time && !end_time->tv_sec && !end_time->tv_nsec) { 470 wait->_qproc = NULL; 471 timed_out = 1; 472 } 473 474 if (end_time && !timed_out) 475 slack = select_estimate_accuracy(end_time); 476 477 retval = 0; 478 for (;;) { 479 unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp; 480 bool can_busy_loop = false; 481 482 inp = fds->in; outp = fds->out; exp = fds->ex; 483 rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex; 484 485 for (i = 0; i < n; ++rinp, ++routp, ++rexp) { 486 unsigned long in, out, ex, all_bits, bit = 1, j; 487 unsigned long res_in = 0, res_out = 0, res_ex = 0; 488 __poll_t mask; 489 490 in = *inp++; out = *outp++; ex = *exp++; 491 all_bits = in | out | ex; 492 if (all_bits == 0) { 493 i += BITS_PER_LONG; 494 continue; 495 } 496 497 for (j = 0; j < BITS_PER_LONG; ++j, ++i, bit <<= 1) { 498 struct fd f; 499 if (i >= n) 500 break; 501 if (!(bit & all_bits)) 502 continue; 503 f = fdget(i); 504 if (f.file) { 505 wait_key_set(wait, in, out, bit, 506 busy_flag); 507 mask = vfs_poll(f.file, wait); 508 509 fdput(f); 510 if ((mask & POLLIN_SET) && (in & bit)) { 511 res_in |= bit; 512 retval++; 513 wait->_qproc = NULL; 514 } 515 if ((mask & POLLOUT_SET) && (out & bit)) { 516 res_out |= bit; 517 retval++; 518 wait->_qproc = NULL; 519 } 520 if ((mask & POLLEX_SET) && (ex & bit)) { 521 res_ex |= bit; 522 retval++; 523 wait->_qproc = NULL; 524 } 525 /* got something, stop busy polling */ 526 if (retval) { 527 can_busy_loop = false; 528 busy_flag = 0; 529 530 /* 531 * only remember a returned 532 * POLL_BUSY_LOOP if we asked for it 533 */ 534 } else if (busy_flag & mask) 535 can_busy_loop = true; 536 537 } 538 } 539 if (res_in) 540 *rinp = res_in; 541 if (res_out) 542 *routp = res_out; 543 if (res_ex) 544 *rexp = res_ex; 545 cond_resched(); 546 } 547 wait->_qproc = NULL; 548 if (retval || timed_out || signal_pending(current)) 549 break; 550 if (table.error) { 551 retval = table.error; 552 break; 553 } 554 555 /* only if found POLL_BUSY_LOOP sockets && not out of time */ 556 if (can_busy_loop && !need_resched()) { 557 if (!busy_start) { 558 busy_start = busy_loop_current_time(); 559 continue; 560 } 561 if (!busy_loop_timeout(busy_start)) 562 continue; 563 } 564 busy_flag = 0; 565 566 /* 567 * If this is the first loop and we have a timeout 568 * given, then we convert to ktime_t and set the to 569 * pointer to the expiry value. 570 */ 571 if (end_time && !to) { 572 expire = timespec64_to_ktime(*end_time); 573 to = &expire; 574 } 575 576 if (!poll_schedule_timeout(&table, TASK_INTERRUPTIBLE, 577 to, slack)) 578 timed_out = 1; 579 } 580 581 poll_freewait(&table); 582 583 return retval; 584 } 585 586 /* 587 * We can actually return ERESTARTSYS instead of EINTR, but I'd 588 * like to be certain this leads to no problems. So I return 589 * EINTR just for safety. 590 * 591 * Update: ERESTARTSYS breaks at least the xview clock binary, so 592 * I'm trying ERESTARTNOHAND which restart only when you want to. 593 */ 594 int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp, 595 fd_set __user *exp, struct timespec64 *end_time) 596 { 597 fd_set_bits fds; 598 void *bits; 599 int ret, max_fds; 600 size_t size, alloc_size; 601 struct fdtable *fdt; 602 /* Allocate small arguments on the stack to save memory and be faster */ 603 long stack_fds[SELECT_STACK_ALLOC/sizeof(long)]; 604 605 ret = -EINVAL; 606 if (n < 0) 607 goto out_nofds; 608 609 /* max_fds can increase, so grab it once to avoid race */ 610 rcu_read_lock(); 611 fdt = files_fdtable(current->files); 612 max_fds = fdt->max_fds; 613 rcu_read_unlock(); 614 if (n > max_fds) 615 n = max_fds; 616 617 /* 618 * We need 6 bitmaps (in/out/ex for both incoming and outgoing), 619 * since we used fdset we need to allocate memory in units of 620 * long-words. 621 */ 622 size = FDS_BYTES(n); 623 bits = stack_fds; 624 if (size > sizeof(stack_fds) / 6) { 625 /* Not enough space in on-stack array; must use kmalloc */ 626 ret = -ENOMEM; 627 if (size > (SIZE_MAX / 6)) 628 goto out_nofds; 629 630 alloc_size = 6 * size; 631 bits = kvmalloc(alloc_size, GFP_KERNEL); 632 if (!bits) 633 goto out_nofds; 634 } 635 fds.in = bits; 636 fds.out = bits + size; 637 fds.ex = bits + 2*size; 638 fds.res_in = bits + 3*size; 639 fds.res_out = bits + 4*size; 640 fds.res_ex = bits + 5*size; 641 642 if ((ret = get_fd_set(n, inp, fds.in)) || 643 (ret = get_fd_set(n, outp, fds.out)) || 644 (ret = get_fd_set(n, exp, fds.ex))) 645 goto out; 646 zero_fd_set(n, fds.res_in); 647 zero_fd_set(n, fds.res_out); 648 zero_fd_set(n, fds.res_ex); 649 650 ret = do_select(n, &fds, end_time); 651 652 if (ret < 0) 653 goto out; 654 if (!ret) { 655 ret = -ERESTARTNOHAND; 656 if (signal_pending(current)) 657 goto out; 658 ret = 0; 659 } 660 661 if (set_fd_set(n, inp, fds.res_in) || 662 set_fd_set(n, outp, fds.res_out) || 663 set_fd_set(n, exp, fds.res_ex)) 664 ret = -EFAULT; 665 666 out: 667 if (bits != stack_fds) 668 kvfree(bits); 669 out_nofds: 670 return ret; 671 } 672 673 static int kern_select(int n, fd_set __user *inp, fd_set __user *outp, 674 fd_set __user *exp, struct timeval __user *tvp) 675 { 676 struct timespec64 end_time, *to = NULL; 677 struct timeval tv; 678 int ret; 679 680 if (tvp) { 681 if (copy_from_user(&tv, tvp, sizeof(tv))) 682 return -EFAULT; 683 684 to = &end_time; 685 if (poll_select_set_timeout(to, 686 tv.tv_sec + (tv.tv_usec / USEC_PER_SEC), 687 (tv.tv_usec % USEC_PER_SEC) * NSEC_PER_USEC)) 688 return -EINVAL; 689 } 690 691 ret = core_sys_select(n, inp, outp, exp, to); 692 ret = poll_select_copy_remaining(&end_time, tvp, 1, ret); 693 694 return ret; 695 } 696 697 SYSCALL_DEFINE5(select, int, n, fd_set __user *, inp, fd_set __user *, outp, 698 fd_set __user *, exp, struct timeval __user *, tvp) 699 { 700 return kern_select(n, inp, outp, exp, tvp); 701 } 702 703 static long do_pselect(int n, fd_set __user *inp, fd_set __user *outp, 704 fd_set __user *exp, struct timespec __user *tsp, 705 const sigset_t __user *sigmask, size_t sigsetsize) 706 { 707 sigset_t ksigmask, sigsaved; 708 struct timespec64 ts, end_time, *to = NULL; 709 int ret; 710 711 if (tsp) { 712 if (get_timespec64(&ts, tsp)) 713 return -EFAULT; 714 715 to = &end_time; 716 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec)) 717 return -EINVAL; 718 } 719 720 if (sigmask) { 721 /* XXX: Don't preclude handling different sized sigset_t's. */ 722 if (sigsetsize != sizeof(sigset_t)) 723 return -EINVAL; 724 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask))) 725 return -EFAULT; 726 727 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP)); 728 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved); 729 } 730 731 ret = core_sys_select(n, inp, outp, exp, to); 732 ret = poll_select_copy_remaining(&end_time, tsp, 0, ret); 733 734 if (ret == -ERESTARTNOHAND) { 735 /* 736 * Don't restore the signal mask yet. Let do_signal() deliver 737 * the signal on the way back to userspace, before the signal 738 * mask is restored. 739 */ 740 if (sigmask) { 741 memcpy(¤t->saved_sigmask, &sigsaved, 742 sizeof(sigsaved)); 743 set_restore_sigmask(); 744 } 745 } else if (sigmask) 746 sigprocmask(SIG_SETMASK, &sigsaved, NULL); 747 748 return ret; 749 } 750 751 /* 752 * Most architectures can't handle 7-argument syscalls. So we provide a 753 * 6-argument version where the sixth argument is a pointer to a structure 754 * which has a pointer to the sigset_t itself followed by a size_t containing 755 * the sigset size. 756 */ 757 SYSCALL_DEFINE6(pselect6, int, n, fd_set __user *, inp, fd_set __user *, outp, 758 fd_set __user *, exp, struct timespec __user *, tsp, 759 void __user *, sig) 760 { 761 size_t sigsetsize = 0; 762 sigset_t __user *up = NULL; 763 764 if (sig) { 765 if (!access_ok(VERIFY_READ, sig, sizeof(void *)+sizeof(size_t)) 766 || __get_user(up, (sigset_t __user * __user *)sig) 767 || __get_user(sigsetsize, 768 (size_t __user *)(sig+sizeof(void *)))) 769 return -EFAULT; 770 } 771 772 return do_pselect(n, inp, outp, exp, tsp, up, sigsetsize); 773 } 774 775 #ifdef __ARCH_WANT_SYS_OLD_SELECT 776 struct sel_arg_struct { 777 unsigned long n; 778 fd_set __user *inp, *outp, *exp; 779 struct timeval __user *tvp; 780 }; 781 782 SYSCALL_DEFINE1(old_select, struct sel_arg_struct __user *, arg) 783 { 784 struct sel_arg_struct a; 785 786 if (copy_from_user(&a, arg, sizeof(a))) 787 return -EFAULT; 788 return kern_select(a.n, a.inp, a.outp, a.exp, a.tvp); 789 } 790 #endif 791 792 struct poll_list { 793 struct poll_list *next; 794 int len; 795 struct pollfd entries[0]; 796 }; 797 798 #define POLLFD_PER_PAGE ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd)) 799 800 /* 801 * Fish for pollable events on the pollfd->fd file descriptor. We're only 802 * interested in events matching the pollfd->events mask, and the result 803 * matching that mask is both recorded in pollfd->revents and returned. The 804 * pwait poll_table will be used by the fd-provided poll handler for waiting, 805 * if pwait->_qproc is non-NULL. 806 */ 807 static inline __poll_t do_pollfd(struct pollfd *pollfd, poll_table *pwait, 808 bool *can_busy_poll, 809 __poll_t busy_flag) 810 { 811 int fd = pollfd->fd; 812 __poll_t mask = 0, filter; 813 struct fd f; 814 815 if (fd < 0) 816 goto out; 817 mask = EPOLLNVAL; 818 f = fdget(fd); 819 if (!f.file) 820 goto out; 821 822 /* userland u16 ->events contains POLL... bitmap */ 823 filter = demangle_poll(pollfd->events) | EPOLLERR | EPOLLHUP; 824 pwait->_key = filter | busy_flag; 825 mask = vfs_poll(f.file, pwait); 826 if (mask & busy_flag) 827 *can_busy_poll = true; 828 mask &= filter; /* Mask out unneeded events. */ 829 fdput(f); 830 831 out: 832 /* ... and so does ->revents */ 833 pollfd->revents = mangle_poll(mask); 834 return mask; 835 } 836 837 static int do_poll(struct poll_list *list, struct poll_wqueues *wait, 838 struct timespec64 *end_time) 839 { 840 poll_table* pt = &wait->pt; 841 ktime_t expire, *to = NULL; 842 int timed_out = 0, count = 0; 843 u64 slack = 0; 844 __poll_t busy_flag = net_busy_loop_on() ? POLL_BUSY_LOOP : 0; 845 unsigned long busy_start = 0; 846 847 /* Optimise the no-wait case */ 848 if (end_time && !end_time->tv_sec && !end_time->tv_nsec) { 849 pt->_qproc = NULL; 850 timed_out = 1; 851 } 852 853 if (end_time && !timed_out) 854 slack = select_estimate_accuracy(end_time); 855 856 for (;;) { 857 struct poll_list *walk; 858 bool can_busy_loop = false; 859 860 for (walk = list; walk != NULL; walk = walk->next) { 861 struct pollfd * pfd, * pfd_end; 862 863 pfd = walk->entries; 864 pfd_end = pfd + walk->len; 865 for (; pfd != pfd_end; pfd++) { 866 /* 867 * Fish for events. If we found one, record it 868 * and kill poll_table->_qproc, so we don't 869 * needlessly register any other waiters after 870 * this. They'll get immediately deregistered 871 * when we break out and return. 872 */ 873 if (do_pollfd(pfd, pt, &can_busy_loop, 874 busy_flag)) { 875 count++; 876 pt->_qproc = NULL; 877 /* found something, stop busy polling */ 878 busy_flag = 0; 879 can_busy_loop = false; 880 } 881 } 882 } 883 /* 884 * All waiters have already been registered, so don't provide 885 * a poll_table->_qproc to them on the next loop iteration. 886 */ 887 pt->_qproc = NULL; 888 if (!count) { 889 count = wait->error; 890 if (signal_pending(current)) 891 count = -EINTR; 892 } 893 if (count || timed_out) 894 break; 895 896 /* only if found POLL_BUSY_LOOP sockets && not out of time */ 897 if (can_busy_loop && !need_resched()) { 898 if (!busy_start) { 899 busy_start = busy_loop_current_time(); 900 continue; 901 } 902 if (!busy_loop_timeout(busy_start)) 903 continue; 904 } 905 busy_flag = 0; 906 907 /* 908 * If this is the first loop and we have a timeout 909 * given, then we convert to ktime_t and set the to 910 * pointer to the expiry value. 911 */ 912 if (end_time && !to) { 913 expire = timespec64_to_ktime(*end_time); 914 to = &expire; 915 } 916 917 if (!poll_schedule_timeout(wait, TASK_INTERRUPTIBLE, to, slack)) 918 timed_out = 1; 919 } 920 return count; 921 } 922 923 #define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list)) / \ 924 sizeof(struct pollfd)) 925 926 static int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds, 927 struct timespec64 *end_time) 928 { 929 struct poll_wqueues table; 930 int err = -EFAULT, fdcount, len, size; 931 /* Allocate small arguments on the stack to save memory and be 932 faster - use long to make sure the buffer is aligned properly 933 on 64 bit archs to avoid unaligned access */ 934 long stack_pps[POLL_STACK_ALLOC/sizeof(long)]; 935 struct poll_list *const head = (struct poll_list *)stack_pps; 936 struct poll_list *walk = head; 937 unsigned long todo = nfds; 938 939 if (nfds > rlimit(RLIMIT_NOFILE)) 940 return -EINVAL; 941 942 len = min_t(unsigned int, nfds, N_STACK_PPS); 943 for (;;) { 944 walk->next = NULL; 945 walk->len = len; 946 if (!len) 947 break; 948 949 if (copy_from_user(walk->entries, ufds + nfds-todo, 950 sizeof(struct pollfd) * walk->len)) 951 goto out_fds; 952 953 todo -= walk->len; 954 if (!todo) 955 break; 956 957 len = min(todo, POLLFD_PER_PAGE); 958 size = sizeof(struct poll_list) + sizeof(struct pollfd) * len; 959 walk = walk->next = kmalloc(size, GFP_KERNEL); 960 if (!walk) { 961 err = -ENOMEM; 962 goto out_fds; 963 } 964 } 965 966 poll_initwait(&table); 967 fdcount = do_poll(head, &table, end_time); 968 poll_freewait(&table); 969 970 for (walk = head; walk; walk = walk->next) { 971 struct pollfd *fds = walk->entries; 972 int j; 973 974 for (j = 0; j < walk->len; j++, ufds++) 975 if (__put_user(fds[j].revents, &ufds->revents)) 976 goto out_fds; 977 } 978 979 err = fdcount; 980 out_fds: 981 walk = head->next; 982 while (walk) { 983 struct poll_list *pos = walk; 984 walk = walk->next; 985 kfree(pos); 986 } 987 988 return err; 989 } 990 991 static long do_restart_poll(struct restart_block *restart_block) 992 { 993 struct pollfd __user *ufds = restart_block->poll.ufds; 994 int nfds = restart_block->poll.nfds; 995 struct timespec64 *to = NULL, end_time; 996 int ret; 997 998 if (restart_block->poll.has_timeout) { 999 end_time.tv_sec = restart_block->poll.tv_sec; 1000 end_time.tv_nsec = restart_block->poll.tv_nsec; 1001 to = &end_time; 1002 } 1003 1004 ret = do_sys_poll(ufds, nfds, to); 1005 1006 if (ret == -EINTR) { 1007 restart_block->fn = do_restart_poll; 1008 ret = -ERESTART_RESTARTBLOCK; 1009 } 1010 return ret; 1011 } 1012 1013 SYSCALL_DEFINE3(poll, struct pollfd __user *, ufds, unsigned int, nfds, 1014 int, timeout_msecs) 1015 { 1016 struct timespec64 end_time, *to = NULL; 1017 int ret; 1018 1019 if (timeout_msecs >= 0) { 1020 to = &end_time; 1021 poll_select_set_timeout(to, timeout_msecs / MSEC_PER_SEC, 1022 NSEC_PER_MSEC * (timeout_msecs % MSEC_PER_SEC)); 1023 } 1024 1025 ret = do_sys_poll(ufds, nfds, to); 1026 1027 if (ret == -EINTR) { 1028 struct restart_block *restart_block; 1029 1030 restart_block = ¤t->restart_block; 1031 restart_block->fn = do_restart_poll; 1032 restart_block->poll.ufds = ufds; 1033 restart_block->poll.nfds = nfds; 1034 1035 if (timeout_msecs >= 0) { 1036 restart_block->poll.tv_sec = end_time.tv_sec; 1037 restart_block->poll.tv_nsec = end_time.tv_nsec; 1038 restart_block->poll.has_timeout = 1; 1039 } else 1040 restart_block->poll.has_timeout = 0; 1041 1042 ret = -ERESTART_RESTARTBLOCK; 1043 } 1044 return ret; 1045 } 1046 1047 SYSCALL_DEFINE5(ppoll, struct pollfd __user *, ufds, unsigned int, nfds, 1048 struct timespec __user *, tsp, const sigset_t __user *, sigmask, 1049 size_t, sigsetsize) 1050 { 1051 sigset_t ksigmask, sigsaved; 1052 struct timespec64 ts, end_time, *to = NULL; 1053 int ret; 1054 1055 if (tsp) { 1056 if (get_timespec64(&ts, tsp)) 1057 return -EFAULT; 1058 1059 to = &end_time; 1060 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec)) 1061 return -EINVAL; 1062 } 1063 1064 if (sigmask) { 1065 /* XXX: Don't preclude handling different sized sigset_t's. */ 1066 if (sigsetsize != sizeof(sigset_t)) 1067 return -EINVAL; 1068 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask))) 1069 return -EFAULT; 1070 1071 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP)); 1072 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved); 1073 } 1074 1075 ret = do_sys_poll(ufds, nfds, to); 1076 1077 /* We can restart this syscall, usually */ 1078 if (ret == -EINTR) { 1079 /* 1080 * Don't restore the signal mask yet. Let do_signal() deliver 1081 * the signal on the way back to userspace, before the signal 1082 * mask is restored. 1083 */ 1084 if (sigmask) { 1085 memcpy(¤t->saved_sigmask, &sigsaved, 1086 sizeof(sigsaved)); 1087 set_restore_sigmask(); 1088 } 1089 ret = -ERESTARTNOHAND; 1090 } else if (sigmask) 1091 sigprocmask(SIG_SETMASK, &sigsaved, NULL); 1092 1093 ret = poll_select_copy_remaining(&end_time, tsp, 0, ret); 1094 1095 return ret; 1096 } 1097 1098 #ifdef CONFIG_COMPAT 1099 #define __COMPAT_NFDBITS (8 * sizeof(compat_ulong_t)) 1100 1101 static 1102 int compat_poll_select_copy_remaining(struct timespec64 *end_time, void __user *p, 1103 int timeval, int ret) 1104 { 1105 struct timespec64 ts; 1106 1107 if (!p) 1108 return ret; 1109 1110 if (current->personality & STICKY_TIMEOUTS) 1111 goto sticky; 1112 1113 /* No update for zero timeout */ 1114 if (!end_time->tv_sec && !end_time->tv_nsec) 1115 return ret; 1116 1117 ktime_get_ts64(&ts); 1118 ts = timespec64_sub(*end_time, ts); 1119 if (ts.tv_sec < 0) 1120 ts.tv_sec = ts.tv_nsec = 0; 1121 1122 if (timeval) { 1123 struct compat_timeval rtv; 1124 1125 rtv.tv_sec = ts.tv_sec; 1126 rtv.tv_usec = ts.tv_nsec / NSEC_PER_USEC; 1127 1128 if (!copy_to_user(p, &rtv, sizeof(rtv))) 1129 return ret; 1130 } else { 1131 if (!compat_put_timespec64(&ts, p)) 1132 return ret; 1133 } 1134 /* 1135 * If an application puts its timeval in read-only memory, we 1136 * don't want the Linux-specific update to the timeval to 1137 * cause a fault after the select has completed 1138 * successfully. However, because we're not updating the 1139 * timeval, we can't restart the system call. 1140 */ 1141 1142 sticky: 1143 if (ret == -ERESTARTNOHAND) 1144 ret = -EINTR; 1145 return ret; 1146 } 1147 1148 /* 1149 * Ooo, nasty. We need here to frob 32-bit unsigned longs to 1150 * 64-bit unsigned longs. 1151 */ 1152 static 1153 int compat_get_fd_set(unsigned long nr, compat_ulong_t __user *ufdset, 1154 unsigned long *fdset) 1155 { 1156 if (ufdset) { 1157 return compat_get_bitmap(fdset, ufdset, nr); 1158 } else { 1159 zero_fd_set(nr, fdset); 1160 return 0; 1161 } 1162 } 1163 1164 static 1165 int compat_set_fd_set(unsigned long nr, compat_ulong_t __user *ufdset, 1166 unsigned long *fdset) 1167 { 1168 if (!ufdset) 1169 return 0; 1170 return compat_put_bitmap(ufdset, fdset, nr); 1171 } 1172 1173 1174 /* 1175 * This is a virtual copy of sys_select from fs/select.c and probably 1176 * should be compared to it from time to time 1177 */ 1178 1179 /* 1180 * We can actually return ERESTARTSYS instead of EINTR, but I'd 1181 * like to be certain this leads to no problems. So I return 1182 * EINTR just for safety. 1183 * 1184 * Update: ERESTARTSYS breaks at least the xview clock binary, so 1185 * I'm trying ERESTARTNOHAND which restart only when you want to. 1186 */ 1187 static int compat_core_sys_select(int n, compat_ulong_t __user *inp, 1188 compat_ulong_t __user *outp, compat_ulong_t __user *exp, 1189 struct timespec64 *end_time) 1190 { 1191 fd_set_bits fds; 1192 void *bits; 1193 int size, max_fds, ret = -EINVAL; 1194 struct fdtable *fdt; 1195 long stack_fds[SELECT_STACK_ALLOC/sizeof(long)]; 1196 1197 if (n < 0) 1198 goto out_nofds; 1199 1200 /* max_fds can increase, so grab it once to avoid race */ 1201 rcu_read_lock(); 1202 fdt = files_fdtable(current->files); 1203 max_fds = fdt->max_fds; 1204 rcu_read_unlock(); 1205 if (n > max_fds) 1206 n = max_fds; 1207 1208 /* 1209 * We need 6 bitmaps (in/out/ex for both incoming and outgoing), 1210 * since we used fdset we need to allocate memory in units of 1211 * long-words. 1212 */ 1213 size = FDS_BYTES(n); 1214 bits = stack_fds; 1215 if (size > sizeof(stack_fds) / 6) { 1216 bits = kmalloc_array(6, size, GFP_KERNEL); 1217 ret = -ENOMEM; 1218 if (!bits) 1219 goto out_nofds; 1220 } 1221 fds.in = (unsigned long *) bits; 1222 fds.out = (unsigned long *) (bits + size); 1223 fds.ex = (unsigned long *) (bits + 2*size); 1224 fds.res_in = (unsigned long *) (bits + 3*size); 1225 fds.res_out = (unsigned long *) (bits + 4*size); 1226 fds.res_ex = (unsigned long *) (bits + 5*size); 1227 1228 if ((ret = compat_get_fd_set(n, inp, fds.in)) || 1229 (ret = compat_get_fd_set(n, outp, fds.out)) || 1230 (ret = compat_get_fd_set(n, exp, fds.ex))) 1231 goto out; 1232 zero_fd_set(n, fds.res_in); 1233 zero_fd_set(n, fds.res_out); 1234 zero_fd_set(n, fds.res_ex); 1235 1236 ret = do_select(n, &fds, end_time); 1237 1238 if (ret < 0) 1239 goto out; 1240 if (!ret) { 1241 ret = -ERESTARTNOHAND; 1242 if (signal_pending(current)) 1243 goto out; 1244 ret = 0; 1245 } 1246 1247 if (compat_set_fd_set(n, inp, fds.res_in) || 1248 compat_set_fd_set(n, outp, fds.res_out) || 1249 compat_set_fd_set(n, exp, fds.res_ex)) 1250 ret = -EFAULT; 1251 out: 1252 if (bits != stack_fds) 1253 kfree(bits); 1254 out_nofds: 1255 return ret; 1256 } 1257 1258 static int do_compat_select(int n, compat_ulong_t __user *inp, 1259 compat_ulong_t __user *outp, compat_ulong_t __user *exp, 1260 struct compat_timeval __user *tvp) 1261 { 1262 struct timespec64 end_time, *to = NULL; 1263 struct compat_timeval tv; 1264 int ret; 1265 1266 if (tvp) { 1267 if (copy_from_user(&tv, tvp, sizeof(tv))) 1268 return -EFAULT; 1269 1270 to = &end_time; 1271 if (poll_select_set_timeout(to, 1272 tv.tv_sec + (tv.tv_usec / USEC_PER_SEC), 1273 (tv.tv_usec % USEC_PER_SEC) * NSEC_PER_USEC)) 1274 return -EINVAL; 1275 } 1276 1277 ret = compat_core_sys_select(n, inp, outp, exp, to); 1278 ret = compat_poll_select_copy_remaining(&end_time, tvp, 1, ret); 1279 1280 return ret; 1281 } 1282 1283 COMPAT_SYSCALL_DEFINE5(select, int, n, compat_ulong_t __user *, inp, 1284 compat_ulong_t __user *, outp, compat_ulong_t __user *, exp, 1285 struct compat_timeval __user *, tvp) 1286 { 1287 return do_compat_select(n, inp, outp, exp, tvp); 1288 } 1289 1290 struct compat_sel_arg_struct { 1291 compat_ulong_t n; 1292 compat_uptr_t inp; 1293 compat_uptr_t outp; 1294 compat_uptr_t exp; 1295 compat_uptr_t tvp; 1296 }; 1297 1298 COMPAT_SYSCALL_DEFINE1(old_select, struct compat_sel_arg_struct __user *, arg) 1299 { 1300 struct compat_sel_arg_struct a; 1301 1302 if (copy_from_user(&a, arg, sizeof(a))) 1303 return -EFAULT; 1304 return do_compat_select(a.n, compat_ptr(a.inp), compat_ptr(a.outp), 1305 compat_ptr(a.exp), compat_ptr(a.tvp)); 1306 } 1307 1308 static long do_compat_pselect(int n, compat_ulong_t __user *inp, 1309 compat_ulong_t __user *outp, compat_ulong_t __user *exp, 1310 struct compat_timespec __user *tsp, compat_sigset_t __user *sigmask, 1311 compat_size_t sigsetsize) 1312 { 1313 sigset_t ksigmask, sigsaved; 1314 struct timespec64 ts, end_time, *to = NULL; 1315 int ret; 1316 1317 if (tsp) { 1318 if (compat_get_timespec64(&ts, tsp)) 1319 return -EFAULT; 1320 1321 to = &end_time; 1322 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec)) 1323 return -EINVAL; 1324 } 1325 1326 if (sigmask) { 1327 if (sigsetsize != sizeof(compat_sigset_t)) 1328 return -EINVAL; 1329 if (get_compat_sigset(&ksigmask, sigmask)) 1330 return -EFAULT; 1331 1332 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP)); 1333 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved); 1334 } 1335 1336 ret = compat_core_sys_select(n, inp, outp, exp, to); 1337 ret = compat_poll_select_copy_remaining(&end_time, tsp, 0, ret); 1338 1339 if (ret == -ERESTARTNOHAND) { 1340 /* 1341 * Don't restore the signal mask yet. Let do_signal() deliver 1342 * the signal on the way back to userspace, before the signal 1343 * mask is restored. 1344 */ 1345 if (sigmask) { 1346 memcpy(¤t->saved_sigmask, &sigsaved, 1347 sizeof(sigsaved)); 1348 set_restore_sigmask(); 1349 } 1350 } else if (sigmask) 1351 sigprocmask(SIG_SETMASK, &sigsaved, NULL); 1352 1353 return ret; 1354 } 1355 1356 COMPAT_SYSCALL_DEFINE6(pselect6, int, n, compat_ulong_t __user *, inp, 1357 compat_ulong_t __user *, outp, compat_ulong_t __user *, exp, 1358 struct compat_timespec __user *, tsp, void __user *, sig) 1359 { 1360 compat_size_t sigsetsize = 0; 1361 compat_uptr_t up = 0; 1362 1363 if (sig) { 1364 if (!access_ok(VERIFY_READ, sig, 1365 sizeof(compat_uptr_t)+sizeof(compat_size_t)) || 1366 __get_user(up, (compat_uptr_t __user *)sig) || 1367 __get_user(sigsetsize, 1368 (compat_size_t __user *)(sig+sizeof(up)))) 1369 return -EFAULT; 1370 } 1371 return do_compat_pselect(n, inp, outp, exp, tsp, compat_ptr(up), 1372 sigsetsize); 1373 } 1374 1375 COMPAT_SYSCALL_DEFINE5(ppoll, struct pollfd __user *, ufds, 1376 unsigned int, nfds, struct compat_timespec __user *, tsp, 1377 const compat_sigset_t __user *, sigmask, compat_size_t, sigsetsize) 1378 { 1379 sigset_t ksigmask, sigsaved; 1380 struct timespec64 ts, end_time, *to = NULL; 1381 int ret; 1382 1383 if (tsp) { 1384 if (compat_get_timespec64(&ts, tsp)) 1385 return -EFAULT; 1386 1387 to = &end_time; 1388 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec)) 1389 return -EINVAL; 1390 } 1391 1392 if (sigmask) { 1393 if (sigsetsize != sizeof(compat_sigset_t)) 1394 return -EINVAL; 1395 if (get_compat_sigset(&ksigmask, sigmask)) 1396 return -EFAULT; 1397 1398 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP)); 1399 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved); 1400 } 1401 1402 ret = do_sys_poll(ufds, nfds, to); 1403 1404 /* We can restart this syscall, usually */ 1405 if (ret == -EINTR) { 1406 /* 1407 * Don't restore the signal mask yet. Let do_signal() deliver 1408 * the signal on the way back to userspace, before the signal 1409 * mask is restored. 1410 */ 1411 if (sigmask) { 1412 memcpy(¤t->saved_sigmask, &sigsaved, 1413 sizeof(sigsaved)); 1414 set_restore_sigmask(); 1415 } 1416 ret = -ERESTARTNOHAND; 1417 } else if (sigmask) 1418 sigprocmask(SIG_SETMASK, &sigsaved, NULL); 1419 1420 ret = compat_poll_select_copy_remaining(&end_time, tsp, 0, ret); 1421 1422 return ret; 1423 } 1424 #endif 1425