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 && !((unsigned long)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 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 EXPORT_SYMBOL(poll_schedule_timeout); 262 263 /** 264 * poll_select_set_timeout - helper function to setup the timeout value 265 * @to: pointer to timespec64 variable for the final timeout 266 * @sec: seconds (from user space) 267 * @nsec: nanoseconds (from user space) 268 * 269 * Note, we do not use a timespec for the user space value here, That 270 * way we can use the function for timeval and compat interfaces as well. 271 * 272 * Returns -EINVAL if sec/nsec are not normalized. Otherwise 0. 273 */ 274 int poll_select_set_timeout(struct timespec64 *to, time64_t sec, long nsec) 275 { 276 struct timespec64 ts = {.tv_sec = sec, .tv_nsec = nsec}; 277 278 if (!timespec64_valid(&ts)) 279 return -EINVAL; 280 281 /* Optimize for the zero timeout value here */ 282 if (!sec && !nsec) { 283 to->tv_sec = to->tv_nsec = 0; 284 } else { 285 ktime_get_ts64(to); 286 *to = timespec64_add_safe(*to, ts); 287 } 288 return 0; 289 } 290 291 static int poll_select_copy_remaining(struct timespec64 *end_time, 292 void __user *p, 293 int timeval, int ret) 294 { 295 struct timespec64 rts; 296 struct timeval rtv; 297 298 if (!p) 299 return ret; 300 301 if (current->personality & STICKY_TIMEOUTS) 302 goto sticky; 303 304 /* No update for zero timeout */ 305 if (!end_time->tv_sec && !end_time->tv_nsec) 306 return ret; 307 308 ktime_get_ts64(&rts); 309 rts = timespec64_sub(*end_time, rts); 310 if (rts.tv_sec < 0) 311 rts.tv_sec = rts.tv_nsec = 0; 312 313 314 if (timeval) { 315 if (sizeof(rtv) > sizeof(rtv.tv_sec) + sizeof(rtv.tv_usec)) 316 memset(&rtv, 0, sizeof(rtv)); 317 rtv.tv_sec = rts.tv_sec; 318 rtv.tv_usec = rts.tv_nsec / NSEC_PER_USEC; 319 320 if (!copy_to_user(p, &rtv, sizeof(rtv))) 321 return ret; 322 323 } else if (!put_timespec64(&rts, p)) 324 return ret; 325 326 /* 327 * If an application puts its timeval in read-only memory, we 328 * don't want the Linux-specific update to the timeval to 329 * cause a fault after the select has completed 330 * successfully. However, because we're not updating the 331 * timeval, we can't restart the system call. 332 */ 333 334 sticky: 335 if (ret == -ERESTARTNOHAND) 336 ret = -EINTR; 337 return ret; 338 } 339 340 /* 341 * Scalable version of the fd_set. 342 */ 343 344 typedef struct { 345 unsigned long *in, *out, *ex; 346 unsigned long *res_in, *res_out, *res_ex; 347 } fd_set_bits; 348 349 /* 350 * How many longwords for "nr" bits? 351 */ 352 #define FDS_BITPERLONG (8*sizeof(long)) 353 #define FDS_LONGS(nr) (((nr)+FDS_BITPERLONG-1)/FDS_BITPERLONG) 354 #define FDS_BYTES(nr) (FDS_LONGS(nr)*sizeof(long)) 355 356 /* 357 * We do a VERIFY_WRITE here even though we are only reading this time: 358 * we'll write to it eventually.. 359 * 360 * Use "unsigned long" accesses to let user-mode fd_set's be long-aligned. 361 */ 362 static inline 363 int get_fd_set(unsigned long nr, void __user *ufdset, unsigned long *fdset) 364 { 365 nr = FDS_BYTES(nr); 366 if (ufdset) 367 return copy_from_user(fdset, ufdset, nr) ? -EFAULT : 0; 368 369 memset(fdset, 0, nr); 370 return 0; 371 } 372 373 static inline unsigned long __must_check 374 set_fd_set(unsigned long nr, void __user *ufdset, unsigned long *fdset) 375 { 376 if (ufdset) 377 return __copy_to_user(ufdset, fdset, FDS_BYTES(nr)); 378 return 0; 379 } 380 381 static inline 382 void zero_fd_set(unsigned long nr, unsigned long *fdset) 383 { 384 memset(fdset, 0, FDS_BYTES(nr)); 385 } 386 387 #define FDS_IN(fds, n) (fds->in + n) 388 #define FDS_OUT(fds, n) (fds->out + n) 389 #define FDS_EX(fds, n) (fds->ex + n) 390 391 #define BITS(fds, n) (*FDS_IN(fds, n)|*FDS_OUT(fds, n)|*FDS_EX(fds, n)) 392 393 static int max_select_fd(unsigned long n, fd_set_bits *fds) 394 { 395 unsigned long *open_fds; 396 unsigned long set; 397 int max; 398 struct fdtable *fdt; 399 400 /* handle last in-complete long-word first */ 401 set = ~(~0UL << (n & (BITS_PER_LONG-1))); 402 n /= BITS_PER_LONG; 403 fdt = files_fdtable(current->files); 404 open_fds = fdt->open_fds + n; 405 max = 0; 406 if (set) { 407 set &= BITS(fds, n); 408 if (set) { 409 if (!(set & ~*open_fds)) 410 goto get_max; 411 return -EBADF; 412 } 413 } 414 while (n) { 415 open_fds--; 416 n--; 417 set = BITS(fds, n); 418 if (!set) 419 continue; 420 if (set & ~*open_fds) 421 return -EBADF; 422 if (max) 423 continue; 424 get_max: 425 do { 426 max++; 427 set >>= 1; 428 } while (set); 429 max += n * BITS_PER_LONG; 430 } 431 432 return max; 433 } 434 435 #define POLLIN_SET (POLLRDNORM | POLLRDBAND | POLLIN | POLLHUP | POLLERR) 436 #define POLLOUT_SET (POLLWRBAND | POLLWRNORM | POLLOUT | POLLERR) 437 #define POLLEX_SET (POLLPRI) 438 439 static inline void wait_key_set(poll_table *wait, unsigned long in, 440 unsigned long out, unsigned long bit, 441 unsigned int ll_flag) 442 { 443 wait->_key = POLLEX_SET | ll_flag; 444 if (in & bit) 445 wait->_key |= POLLIN_SET; 446 if (out & bit) 447 wait->_key |= POLLOUT_SET; 448 } 449 450 static int do_select(int n, fd_set_bits *fds, struct timespec64 *end_time) 451 { 452 ktime_t expire, *to = NULL; 453 struct poll_wqueues table; 454 poll_table *wait; 455 int retval, i, timed_out = 0; 456 u64 slack = 0; 457 unsigned int busy_flag = net_busy_loop_on() ? POLL_BUSY_LOOP : 0; 458 unsigned long busy_start = 0; 459 460 rcu_read_lock(); 461 retval = max_select_fd(n, fds); 462 rcu_read_unlock(); 463 464 if (retval < 0) 465 return retval; 466 n = retval; 467 468 poll_initwait(&table); 469 wait = &table.pt; 470 if (end_time && !end_time->tv_sec && !end_time->tv_nsec) { 471 wait->_qproc = NULL; 472 timed_out = 1; 473 } 474 475 if (end_time && !timed_out) 476 slack = select_estimate_accuracy(end_time); 477 478 retval = 0; 479 for (;;) { 480 unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp; 481 bool can_busy_loop = false; 482 483 inp = fds->in; outp = fds->out; exp = fds->ex; 484 rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex; 485 486 for (i = 0; i < n; ++rinp, ++routp, ++rexp) { 487 unsigned long in, out, ex, all_bits, bit = 1, mask, j; 488 unsigned long res_in = 0, res_out = 0, res_ex = 0; 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 const struct file_operations *f_op; 506 f_op = f.file->f_op; 507 mask = DEFAULT_POLLMASK; 508 if (f_op->poll) { 509 wait_key_set(wait, in, out, 510 bit, busy_flag); 511 mask = (*f_op->poll)(f.file, wait); 512 } 513 fdput(f); 514 if ((mask & POLLIN_SET) && (in & bit)) { 515 res_in |= bit; 516 retval++; 517 wait->_qproc = NULL; 518 } 519 if ((mask & POLLOUT_SET) && (out & bit)) { 520 res_out |= bit; 521 retval++; 522 wait->_qproc = NULL; 523 } 524 if ((mask & POLLEX_SET) && (ex & bit)) { 525 res_ex |= bit; 526 retval++; 527 wait->_qproc = NULL; 528 } 529 /* got something, stop busy polling */ 530 if (retval) { 531 can_busy_loop = false; 532 busy_flag = 0; 533 534 /* 535 * only remember a returned 536 * POLL_BUSY_LOOP if we asked for it 537 */ 538 } else if (busy_flag & mask) 539 can_busy_loop = true; 540 541 } 542 } 543 if (res_in) 544 *rinp = res_in; 545 if (res_out) 546 *routp = res_out; 547 if (res_ex) 548 *rexp = res_ex; 549 cond_resched(); 550 } 551 wait->_qproc = NULL; 552 if (retval || timed_out || signal_pending(current)) 553 break; 554 if (table.error) { 555 retval = table.error; 556 break; 557 } 558 559 /* only if found POLL_BUSY_LOOP sockets && not out of time */ 560 if (can_busy_loop && !need_resched()) { 561 if (!busy_start) { 562 busy_start = busy_loop_current_time(); 563 continue; 564 } 565 if (!busy_loop_timeout(busy_start)) 566 continue; 567 } 568 busy_flag = 0; 569 570 /* 571 * If this is the first loop and we have a timeout 572 * given, then we convert to ktime_t and set the to 573 * pointer to the expiry value. 574 */ 575 if (end_time && !to) { 576 expire = timespec64_to_ktime(*end_time); 577 to = &expire; 578 } 579 580 if (!poll_schedule_timeout(&table, TASK_INTERRUPTIBLE, 581 to, slack)) 582 timed_out = 1; 583 } 584 585 poll_freewait(&table); 586 587 return retval; 588 } 589 590 /* 591 * We can actually return ERESTARTSYS instead of EINTR, but I'd 592 * like to be certain this leads to no problems. So I return 593 * EINTR just for safety. 594 * 595 * Update: ERESTARTSYS breaks at least the xview clock binary, so 596 * I'm trying ERESTARTNOHAND which restart only when you want to. 597 */ 598 int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp, 599 fd_set __user *exp, struct timespec64 *end_time) 600 { 601 fd_set_bits fds; 602 void *bits; 603 int ret, max_fds; 604 size_t size, alloc_size; 605 struct fdtable *fdt; 606 /* Allocate small arguments on the stack to save memory and be faster */ 607 long stack_fds[SELECT_STACK_ALLOC/sizeof(long)]; 608 609 ret = -EINVAL; 610 if (n < 0) 611 goto out_nofds; 612 613 /* max_fds can increase, so grab it once to avoid race */ 614 rcu_read_lock(); 615 fdt = files_fdtable(current->files); 616 max_fds = fdt->max_fds; 617 rcu_read_unlock(); 618 if (n > max_fds) 619 n = max_fds; 620 621 /* 622 * We need 6 bitmaps (in/out/ex for both incoming and outgoing), 623 * since we used fdset we need to allocate memory in units of 624 * long-words. 625 */ 626 size = FDS_BYTES(n); 627 bits = stack_fds; 628 if (size > sizeof(stack_fds) / 6) { 629 /* Not enough space in on-stack array; must use kmalloc */ 630 ret = -ENOMEM; 631 if (size > (SIZE_MAX / 6)) 632 goto out_nofds; 633 634 alloc_size = 6 * size; 635 bits = kvmalloc(alloc_size, GFP_KERNEL); 636 if (!bits) 637 goto out_nofds; 638 } 639 fds.in = bits; 640 fds.out = bits + size; 641 fds.ex = bits + 2*size; 642 fds.res_in = bits + 3*size; 643 fds.res_out = bits + 4*size; 644 fds.res_ex = bits + 5*size; 645 646 if ((ret = get_fd_set(n, inp, fds.in)) || 647 (ret = get_fd_set(n, outp, fds.out)) || 648 (ret = get_fd_set(n, exp, fds.ex))) 649 goto out; 650 zero_fd_set(n, fds.res_in); 651 zero_fd_set(n, fds.res_out); 652 zero_fd_set(n, fds.res_ex); 653 654 ret = do_select(n, &fds, end_time); 655 656 if (ret < 0) 657 goto out; 658 if (!ret) { 659 ret = -ERESTARTNOHAND; 660 if (signal_pending(current)) 661 goto out; 662 ret = 0; 663 } 664 665 if (set_fd_set(n, inp, fds.res_in) || 666 set_fd_set(n, outp, fds.res_out) || 667 set_fd_set(n, exp, fds.res_ex)) 668 ret = -EFAULT; 669 670 out: 671 if (bits != stack_fds) 672 kvfree(bits); 673 out_nofds: 674 return ret; 675 } 676 677 SYSCALL_DEFINE5(select, int, n, fd_set __user *, inp, fd_set __user *, outp, 678 fd_set __user *, exp, struct timeval __user *, tvp) 679 { 680 struct timespec64 end_time, *to = NULL; 681 struct timeval tv; 682 int ret; 683 684 if (tvp) { 685 if (copy_from_user(&tv, tvp, sizeof(tv))) 686 return -EFAULT; 687 688 to = &end_time; 689 if (poll_select_set_timeout(to, 690 tv.tv_sec + (tv.tv_usec / USEC_PER_SEC), 691 (tv.tv_usec % USEC_PER_SEC) * NSEC_PER_USEC)) 692 return -EINVAL; 693 } 694 695 ret = core_sys_select(n, inp, outp, exp, to); 696 ret = poll_select_copy_remaining(&end_time, tvp, 1, ret); 697 698 return ret; 699 } 700 701 static long do_pselect(int n, fd_set __user *inp, fd_set __user *outp, 702 fd_set __user *exp, struct timespec __user *tsp, 703 const sigset_t __user *sigmask, size_t sigsetsize) 704 { 705 sigset_t ksigmask, sigsaved; 706 struct timespec64 ts, end_time, *to = NULL; 707 int ret; 708 709 if (tsp) { 710 if (get_timespec64(&ts, tsp)) 711 return -EFAULT; 712 713 to = &end_time; 714 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec)) 715 return -EINVAL; 716 } 717 718 if (sigmask) { 719 /* XXX: Don't preclude handling different sized sigset_t's. */ 720 if (sigsetsize != sizeof(sigset_t)) 721 return -EINVAL; 722 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask))) 723 return -EFAULT; 724 725 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP)); 726 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved); 727 } 728 729 ret = core_sys_select(n, inp, outp, exp, to); 730 ret = poll_select_copy_remaining(&end_time, tsp, 0, ret); 731 732 if (ret == -ERESTARTNOHAND) { 733 /* 734 * Don't restore the signal mask yet. Let do_signal() deliver 735 * the signal on the way back to userspace, before the signal 736 * mask is restored. 737 */ 738 if (sigmask) { 739 memcpy(¤t->saved_sigmask, &sigsaved, 740 sizeof(sigsaved)); 741 set_restore_sigmask(); 742 } 743 } else if (sigmask) 744 sigprocmask(SIG_SETMASK, &sigsaved, NULL); 745 746 return ret; 747 } 748 749 /* 750 * Most architectures can't handle 7-argument syscalls. So we provide a 751 * 6-argument version where the sixth argument is a pointer to a structure 752 * which has a pointer to the sigset_t itself followed by a size_t containing 753 * the sigset size. 754 */ 755 SYSCALL_DEFINE6(pselect6, int, n, fd_set __user *, inp, fd_set __user *, outp, 756 fd_set __user *, exp, struct timespec __user *, tsp, 757 void __user *, sig) 758 { 759 size_t sigsetsize = 0; 760 sigset_t __user *up = NULL; 761 762 if (sig) { 763 if (!access_ok(VERIFY_READ, sig, sizeof(void *)+sizeof(size_t)) 764 || __get_user(up, (sigset_t __user * __user *)sig) 765 || __get_user(sigsetsize, 766 (size_t __user *)(sig+sizeof(void *)))) 767 return -EFAULT; 768 } 769 770 return do_pselect(n, inp, outp, exp, tsp, up, sigsetsize); 771 } 772 773 #ifdef __ARCH_WANT_SYS_OLD_SELECT 774 struct sel_arg_struct { 775 unsigned long n; 776 fd_set __user *inp, *outp, *exp; 777 struct timeval __user *tvp; 778 }; 779 780 SYSCALL_DEFINE1(old_select, struct sel_arg_struct __user *, arg) 781 { 782 struct sel_arg_struct a; 783 784 if (copy_from_user(&a, arg, sizeof(a))) 785 return -EFAULT; 786 return sys_select(a.n, a.inp, a.outp, a.exp, a.tvp); 787 } 788 #endif 789 790 struct poll_list { 791 struct poll_list *next; 792 int len; 793 struct pollfd entries[0]; 794 }; 795 796 #define POLLFD_PER_PAGE ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd)) 797 798 /* 799 * Fish for pollable events on the pollfd->fd file descriptor. We're only 800 * interested in events matching the pollfd->events mask, and the result 801 * matching that mask is both recorded in pollfd->revents and returned. The 802 * pwait poll_table will be used by the fd-provided poll handler for waiting, 803 * if pwait->_qproc is non-NULL. 804 */ 805 static inline unsigned int do_pollfd(struct pollfd *pollfd, poll_table *pwait, 806 bool *can_busy_poll, 807 unsigned int busy_flag) 808 { 809 unsigned int mask; 810 int fd; 811 812 mask = 0; 813 fd = pollfd->fd; 814 if (fd >= 0) { 815 struct fd f = fdget(fd); 816 mask = POLLNVAL; 817 if (f.file) { 818 mask = DEFAULT_POLLMASK; 819 if (f.file->f_op->poll) { 820 pwait->_key = pollfd->events|POLLERR|POLLHUP; 821 pwait->_key |= busy_flag; 822 mask = f.file->f_op->poll(f.file, pwait); 823 if (mask & busy_flag) 824 *can_busy_poll = true; 825 } 826 /* Mask out unneeded events. */ 827 mask &= pollfd->events | POLLERR | POLLHUP; 828 fdput(f); 829 } 830 } 831 pollfd->revents = mask; 832 833 return mask; 834 } 835 836 static int do_poll(struct poll_list *list, struct poll_wqueues *wait, 837 struct timespec64 *end_time) 838 { 839 poll_table* pt = &wait->pt; 840 ktime_t expire, *to = NULL; 841 int timed_out = 0, count = 0; 842 u64 slack = 0; 843 unsigned int busy_flag = net_busy_loop_on() ? POLL_BUSY_LOOP : 0; 844 unsigned long busy_start = 0; 845 846 /* Optimise the no-wait case */ 847 if (end_time && !end_time->tv_sec && !end_time->tv_nsec) { 848 pt->_qproc = NULL; 849 timed_out = 1; 850 } 851 852 if (end_time && !timed_out) 853 slack = select_estimate_accuracy(end_time); 854 855 for (;;) { 856 struct poll_list *walk; 857 bool can_busy_loop = false; 858 859 for (walk = list; walk != NULL; walk = walk->next) { 860 struct pollfd * pfd, * pfd_end; 861 862 pfd = walk->entries; 863 pfd_end = pfd + walk->len; 864 for (; pfd != pfd_end; pfd++) { 865 /* 866 * Fish for events. If we found one, record it 867 * and kill poll_table->_qproc, so we don't 868 * needlessly register any other waiters after 869 * this. They'll get immediately deregistered 870 * when we break out and return. 871 */ 872 if (do_pollfd(pfd, pt, &can_busy_loop, 873 busy_flag)) { 874 count++; 875 pt->_qproc = NULL; 876 /* found something, stop busy polling */ 877 busy_flag = 0; 878 can_busy_loop = false; 879 } 880 } 881 } 882 /* 883 * All waiters have already been registered, so don't provide 884 * a poll_table->_qproc to them on the next loop iteration. 885 */ 886 pt->_qproc = NULL; 887 if (!count) { 888 count = wait->error; 889 if (signal_pending(current)) 890 count = -EINTR; 891 } 892 if (count || timed_out) 893 break; 894 895 /* only if found POLL_BUSY_LOOP sockets && not out of time */ 896 if (can_busy_loop && !need_resched()) { 897 if (!busy_start) { 898 busy_start = busy_loop_current_time(); 899 continue; 900 } 901 if (!busy_loop_timeout(busy_start)) 902 continue; 903 } 904 busy_flag = 0; 905 906 /* 907 * If this is the first loop and we have a timeout 908 * given, then we convert to ktime_t and set the to 909 * pointer to the expiry value. 910 */ 911 if (end_time && !to) { 912 expire = timespec64_to_ktime(*end_time); 913 to = &expire; 914 } 915 916 if (!poll_schedule_timeout(wait, TASK_INTERRUPTIBLE, to, slack)) 917 timed_out = 1; 918 } 919 return count; 920 } 921 922 #define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list)) / \ 923 sizeof(struct pollfd)) 924 925 static int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds, 926 struct timespec64 *end_time) 927 { 928 struct poll_wqueues table; 929 int err = -EFAULT, fdcount, len, size; 930 /* Allocate small arguments on the stack to save memory and be 931 faster - use long to make sure the buffer is aligned properly 932 on 64 bit archs to avoid unaligned access */ 933 long stack_pps[POLL_STACK_ALLOC/sizeof(long)]; 934 struct poll_list *const head = (struct poll_list *)stack_pps; 935 struct poll_list *walk = head; 936 unsigned long todo = nfds; 937 938 if (nfds > rlimit(RLIMIT_NOFILE)) 939 return -EINVAL; 940 941 len = min_t(unsigned int, nfds, N_STACK_PPS); 942 for (;;) { 943 walk->next = NULL; 944 walk->len = len; 945 if (!len) 946 break; 947 948 if (copy_from_user(walk->entries, ufds + nfds-todo, 949 sizeof(struct pollfd) * walk->len)) 950 goto out_fds; 951 952 todo -= walk->len; 953 if (!todo) 954 break; 955 956 len = min(todo, POLLFD_PER_PAGE); 957 size = sizeof(struct poll_list) + sizeof(struct pollfd) * len; 958 walk = walk->next = kmalloc(size, GFP_KERNEL); 959 if (!walk) { 960 err = -ENOMEM; 961 goto out_fds; 962 } 963 } 964 965 poll_initwait(&table); 966 fdcount = do_poll(head, &table, end_time); 967 poll_freewait(&table); 968 969 for (walk = head; walk; walk = walk->next) { 970 struct pollfd *fds = walk->entries; 971 int j; 972 973 for (j = 0; j < walk->len; j++, ufds++) 974 if (__put_user(fds[j].revents, &ufds->revents)) 975 goto out_fds; 976 } 977 978 err = fdcount; 979 out_fds: 980 walk = head->next; 981 while (walk) { 982 struct poll_list *pos = walk; 983 walk = walk->next; 984 kfree(pos); 985 } 986 987 return err; 988 } 989 990 static long do_restart_poll(struct restart_block *restart_block) 991 { 992 struct pollfd __user *ufds = restart_block->poll.ufds; 993 int nfds = restart_block->poll.nfds; 994 struct timespec64 *to = NULL, end_time; 995 int ret; 996 997 if (restart_block->poll.has_timeout) { 998 end_time.tv_sec = restart_block->poll.tv_sec; 999 end_time.tv_nsec = restart_block->poll.tv_nsec; 1000 to = &end_time; 1001 } 1002 1003 ret = do_sys_poll(ufds, nfds, to); 1004 1005 if (ret == -EINTR) { 1006 restart_block->fn = do_restart_poll; 1007 ret = -ERESTART_RESTARTBLOCK; 1008 } 1009 return ret; 1010 } 1011 1012 SYSCALL_DEFINE3(poll, struct pollfd __user *, ufds, unsigned int, nfds, 1013 int, timeout_msecs) 1014 { 1015 struct timespec64 end_time, *to = NULL; 1016 int ret; 1017 1018 if (timeout_msecs >= 0) { 1019 to = &end_time; 1020 poll_select_set_timeout(to, timeout_msecs / MSEC_PER_SEC, 1021 NSEC_PER_MSEC * (timeout_msecs % MSEC_PER_SEC)); 1022 } 1023 1024 ret = do_sys_poll(ufds, nfds, to); 1025 1026 if (ret == -EINTR) { 1027 struct restart_block *restart_block; 1028 1029 restart_block = ¤t->restart_block; 1030 restart_block->fn = do_restart_poll; 1031 restart_block->poll.ufds = ufds; 1032 restart_block->poll.nfds = nfds; 1033 1034 if (timeout_msecs >= 0) { 1035 restart_block->poll.tv_sec = end_time.tv_sec; 1036 restart_block->poll.tv_nsec = end_time.tv_nsec; 1037 restart_block->poll.has_timeout = 1; 1038 } else 1039 restart_block->poll.has_timeout = 0; 1040 1041 ret = -ERESTART_RESTARTBLOCK; 1042 } 1043 return ret; 1044 } 1045 1046 SYSCALL_DEFINE5(ppoll, struct pollfd __user *, ufds, unsigned int, nfds, 1047 struct timespec __user *, tsp, const sigset_t __user *, sigmask, 1048 size_t, sigsetsize) 1049 { 1050 sigset_t ksigmask, sigsaved; 1051 struct timespec64 ts, end_time, *to = NULL; 1052 int ret; 1053 1054 if (tsp) { 1055 if (get_timespec64(&ts, tsp)) 1056 return -EFAULT; 1057 1058 to = &end_time; 1059 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec)) 1060 return -EINVAL; 1061 } 1062 1063 if (sigmask) { 1064 /* XXX: Don't preclude handling different sized sigset_t's. */ 1065 if (sigsetsize != sizeof(sigset_t)) 1066 return -EINVAL; 1067 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask))) 1068 return -EFAULT; 1069 1070 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP)); 1071 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved); 1072 } 1073 1074 ret = do_sys_poll(ufds, nfds, to); 1075 1076 /* We can restart this syscall, usually */ 1077 if (ret == -EINTR) { 1078 /* 1079 * Don't restore the signal mask yet. Let do_signal() deliver 1080 * the signal on the way back to userspace, before the signal 1081 * mask is restored. 1082 */ 1083 if (sigmask) { 1084 memcpy(¤t->saved_sigmask, &sigsaved, 1085 sizeof(sigsaved)); 1086 set_restore_sigmask(); 1087 } 1088 ret = -ERESTARTNOHAND; 1089 } else if (sigmask) 1090 sigprocmask(SIG_SETMASK, &sigsaved, NULL); 1091 1092 ret = poll_select_copy_remaining(&end_time, tsp, 0, ret); 1093 1094 return ret; 1095 } 1096 1097 #ifdef CONFIG_COMPAT 1098 #define __COMPAT_NFDBITS (8 * sizeof(compat_ulong_t)) 1099 1100 static 1101 int compat_poll_select_copy_remaining(struct timespec64 *end_time, void __user *p, 1102 int timeval, int ret) 1103 { 1104 struct timespec64 ts; 1105 1106 if (!p) 1107 return ret; 1108 1109 if (current->personality & STICKY_TIMEOUTS) 1110 goto sticky; 1111 1112 /* No update for zero timeout */ 1113 if (!end_time->tv_sec && !end_time->tv_nsec) 1114 return ret; 1115 1116 ktime_get_ts64(&ts); 1117 ts = timespec64_sub(*end_time, ts); 1118 if (ts.tv_sec < 0) 1119 ts.tv_sec = ts.tv_nsec = 0; 1120 1121 if (timeval) { 1122 struct compat_timeval rtv; 1123 1124 rtv.tv_sec = ts.tv_sec; 1125 rtv.tv_usec = ts.tv_nsec / NSEC_PER_USEC; 1126 1127 if (!copy_to_user(p, &rtv, sizeof(rtv))) 1128 return ret; 1129 } else { 1130 if (!compat_put_timespec64(&ts, p)) 1131 return ret; 1132 } 1133 /* 1134 * If an application puts its timeval in read-only memory, we 1135 * don't want the Linux-specific update to the timeval to 1136 * cause a fault after the select has completed 1137 * successfully. However, because we're not updating the 1138 * timeval, we can't restart the system call. 1139 */ 1140 1141 sticky: 1142 if (ret == -ERESTARTNOHAND) 1143 ret = -EINTR; 1144 return ret; 1145 } 1146 1147 /* 1148 * Ooo, nasty. We need here to frob 32-bit unsigned longs to 1149 * 64-bit unsigned longs. 1150 */ 1151 static 1152 int compat_get_fd_set(unsigned long nr, compat_ulong_t __user *ufdset, 1153 unsigned long *fdset) 1154 { 1155 if (ufdset) { 1156 return compat_get_bitmap(fdset, ufdset, nr); 1157 } else { 1158 zero_fd_set(nr, fdset); 1159 return 0; 1160 } 1161 } 1162 1163 static 1164 int compat_set_fd_set(unsigned long nr, compat_ulong_t __user *ufdset, 1165 unsigned long *fdset) 1166 { 1167 if (!ufdset) 1168 return 0; 1169 return compat_put_bitmap(ufdset, fdset, nr); 1170 } 1171 1172 1173 /* 1174 * This is a virtual copy of sys_select from fs/select.c and probably 1175 * should be compared to it from time to time 1176 */ 1177 1178 /* 1179 * We can actually return ERESTARTSYS instead of EINTR, but I'd 1180 * like to be certain this leads to no problems. So I return 1181 * EINTR just for safety. 1182 * 1183 * Update: ERESTARTSYS breaks at least the xview clock binary, so 1184 * I'm trying ERESTARTNOHAND which restart only when you want to. 1185 */ 1186 static int compat_core_sys_select(int n, compat_ulong_t __user *inp, 1187 compat_ulong_t __user *outp, compat_ulong_t __user *exp, 1188 struct timespec64 *end_time) 1189 { 1190 fd_set_bits fds; 1191 void *bits; 1192 int size, max_fds, ret = -EINVAL; 1193 struct fdtable *fdt; 1194 long stack_fds[SELECT_STACK_ALLOC/sizeof(long)]; 1195 1196 if (n < 0) 1197 goto out_nofds; 1198 1199 /* max_fds can increase, so grab it once to avoid race */ 1200 rcu_read_lock(); 1201 fdt = files_fdtable(current->files); 1202 max_fds = fdt->max_fds; 1203 rcu_read_unlock(); 1204 if (n > max_fds) 1205 n = max_fds; 1206 1207 /* 1208 * We need 6 bitmaps (in/out/ex for both incoming and outgoing), 1209 * since we used fdset we need to allocate memory in units of 1210 * long-words. 1211 */ 1212 size = FDS_BYTES(n); 1213 bits = stack_fds; 1214 if (size > sizeof(stack_fds) / 6) { 1215 bits = kmalloc(6 * size, GFP_KERNEL); 1216 ret = -ENOMEM; 1217 if (!bits) 1218 goto out_nofds; 1219 } 1220 fds.in = (unsigned long *) bits; 1221 fds.out = (unsigned long *) (bits + size); 1222 fds.ex = (unsigned long *) (bits + 2*size); 1223 fds.res_in = (unsigned long *) (bits + 3*size); 1224 fds.res_out = (unsigned long *) (bits + 4*size); 1225 fds.res_ex = (unsigned long *) (bits + 5*size); 1226 1227 if ((ret = compat_get_fd_set(n, inp, fds.in)) || 1228 (ret = compat_get_fd_set(n, outp, fds.out)) || 1229 (ret = compat_get_fd_set(n, exp, fds.ex))) 1230 goto out; 1231 zero_fd_set(n, fds.res_in); 1232 zero_fd_set(n, fds.res_out); 1233 zero_fd_set(n, fds.res_ex); 1234 1235 ret = do_select(n, &fds, end_time); 1236 1237 if (ret < 0) 1238 goto out; 1239 if (!ret) { 1240 ret = -ERESTARTNOHAND; 1241 if (signal_pending(current)) 1242 goto out; 1243 ret = 0; 1244 } 1245 1246 if (compat_set_fd_set(n, inp, fds.res_in) || 1247 compat_set_fd_set(n, outp, fds.res_out) || 1248 compat_set_fd_set(n, exp, fds.res_ex)) 1249 ret = -EFAULT; 1250 out: 1251 if (bits != stack_fds) 1252 kfree(bits); 1253 out_nofds: 1254 return ret; 1255 } 1256 1257 COMPAT_SYSCALL_DEFINE5(select, int, n, compat_ulong_t __user *, inp, 1258 compat_ulong_t __user *, outp, compat_ulong_t __user *, exp, 1259 struct compat_timeval __user *, tvp) 1260 { 1261 struct timespec64 end_time, *to = NULL; 1262 struct compat_timeval tv; 1263 int ret; 1264 1265 if (tvp) { 1266 if (copy_from_user(&tv, tvp, sizeof(tv))) 1267 return -EFAULT; 1268 1269 to = &end_time; 1270 if (poll_select_set_timeout(to, 1271 tv.tv_sec + (tv.tv_usec / USEC_PER_SEC), 1272 (tv.tv_usec % USEC_PER_SEC) * NSEC_PER_USEC)) 1273 return -EINVAL; 1274 } 1275 1276 ret = compat_core_sys_select(n, inp, outp, exp, to); 1277 ret = compat_poll_select_copy_remaining(&end_time, tvp, 1, ret); 1278 1279 return ret; 1280 } 1281 1282 struct compat_sel_arg_struct { 1283 compat_ulong_t n; 1284 compat_uptr_t inp; 1285 compat_uptr_t outp; 1286 compat_uptr_t exp; 1287 compat_uptr_t tvp; 1288 }; 1289 1290 COMPAT_SYSCALL_DEFINE1(old_select, struct compat_sel_arg_struct __user *, arg) 1291 { 1292 struct compat_sel_arg_struct a; 1293 1294 if (copy_from_user(&a, arg, sizeof(a))) 1295 return -EFAULT; 1296 return compat_sys_select(a.n, compat_ptr(a.inp), compat_ptr(a.outp), 1297 compat_ptr(a.exp), compat_ptr(a.tvp)); 1298 } 1299 1300 static long do_compat_pselect(int n, compat_ulong_t __user *inp, 1301 compat_ulong_t __user *outp, compat_ulong_t __user *exp, 1302 struct compat_timespec __user *tsp, compat_sigset_t __user *sigmask, 1303 compat_size_t sigsetsize) 1304 { 1305 sigset_t ksigmask, sigsaved; 1306 struct timespec64 ts, end_time, *to = NULL; 1307 int ret; 1308 1309 if (tsp) { 1310 if (compat_get_timespec64(&ts, tsp)) 1311 return -EFAULT; 1312 1313 to = &end_time; 1314 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec)) 1315 return -EINVAL; 1316 } 1317 1318 if (sigmask) { 1319 if (sigsetsize != sizeof(compat_sigset_t)) 1320 return -EINVAL; 1321 if (get_compat_sigset(&ksigmask, sigmask)) 1322 return -EFAULT; 1323 1324 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP)); 1325 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved); 1326 } 1327 1328 ret = compat_core_sys_select(n, inp, outp, exp, to); 1329 ret = compat_poll_select_copy_remaining(&end_time, tsp, 0, ret); 1330 1331 if (ret == -ERESTARTNOHAND) { 1332 /* 1333 * Don't restore the signal mask yet. Let do_signal() deliver 1334 * the signal on the way back to userspace, before the signal 1335 * mask is restored. 1336 */ 1337 if (sigmask) { 1338 memcpy(¤t->saved_sigmask, &sigsaved, 1339 sizeof(sigsaved)); 1340 set_restore_sigmask(); 1341 } 1342 } else if (sigmask) 1343 sigprocmask(SIG_SETMASK, &sigsaved, NULL); 1344 1345 return ret; 1346 } 1347 1348 COMPAT_SYSCALL_DEFINE6(pselect6, int, n, compat_ulong_t __user *, inp, 1349 compat_ulong_t __user *, outp, compat_ulong_t __user *, exp, 1350 struct compat_timespec __user *, tsp, void __user *, sig) 1351 { 1352 compat_size_t sigsetsize = 0; 1353 compat_uptr_t up = 0; 1354 1355 if (sig) { 1356 if (!access_ok(VERIFY_READ, sig, 1357 sizeof(compat_uptr_t)+sizeof(compat_size_t)) || 1358 __get_user(up, (compat_uptr_t __user *)sig) || 1359 __get_user(sigsetsize, 1360 (compat_size_t __user *)(sig+sizeof(up)))) 1361 return -EFAULT; 1362 } 1363 return do_compat_pselect(n, inp, outp, exp, tsp, compat_ptr(up), 1364 sigsetsize); 1365 } 1366 1367 COMPAT_SYSCALL_DEFINE5(ppoll, struct pollfd __user *, ufds, 1368 unsigned int, nfds, struct compat_timespec __user *, tsp, 1369 const compat_sigset_t __user *, sigmask, compat_size_t, sigsetsize) 1370 { 1371 sigset_t ksigmask, sigsaved; 1372 struct timespec64 ts, end_time, *to = NULL; 1373 int ret; 1374 1375 if (tsp) { 1376 if (compat_get_timespec64(&ts, tsp)) 1377 return -EFAULT; 1378 1379 to = &end_time; 1380 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec)) 1381 return -EINVAL; 1382 } 1383 1384 if (sigmask) { 1385 if (sigsetsize != sizeof(compat_sigset_t)) 1386 return -EINVAL; 1387 if (get_compat_sigset(&ksigmask, sigmask)) 1388 return -EFAULT; 1389 1390 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP)); 1391 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved); 1392 } 1393 1394 ret = do_sys_poll(ufds, nfds, to); 1395 1396 /* We can restart this syscall, usually */ 1397 if (ret == -EINTR) { 1398 /* 1399 * Don't restore the signal mask yet. Let do_signal() deliver 1400 * the signal on the way back to userspace, before the signal 1401 * mask is restored. 1402 */ 1403 if (sigmask) { 1404 memcpy(¤t->saved_sigmask, &sigsaved, 1405 sizeof(sigsaved)); 1406 set_restore_sigmask(); 1407 } 1408 ret = -ERESTARTNOHAND; 1409 } else if (sigmask) 1410 sigprocmask(SIG_SETMASK, &sigsaved, NULL); 1411 1412 ret = compat_poll_select_copy_remaining(&end_time, tsp, 0, ret); 1413 1414 return ret; 1415 } 1416 #endif 1417