1 /* 2 * Copyright (c) 2005-2018 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Jeffrey Hsu and Matthew Dillon. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * 35 * Copyright (c) 1982, 1986, 1989, 1991, 1993 36 * The Regents of the University of California. All rights reserved. 37 * (c) UNIX System Laboratories, Inc. 38 * All or some portions of this file are derived from material licensed 39 * to the University of California by American Telephone and Telegraph 40 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 41 * the permission of UNIX System Laboratories, Inc. 42 * 43 * Redistribution and use in source and binary forms, with or without 44 * modification, are permitted provided that the following conditions 45 * are met: 46 * 1. Redistributions of source code must retain the above copyright 47 * notice, this list of conditions and the following disclaimer. 48 * 2. Redistributions in binary form must reproduce the above copyright 49 * notice, this list of conditions and the following disclaimer in the 50 * documentation and/or other materials provided with the distribution. 51 * 3. Neither the name of the University nor the names of its contributors 52 * may be used to endorse or promote products derived from this software 53 * without specific prior written permission. 54 * 55 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 56 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 57 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 58 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 59 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 60 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 61 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 62 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 63 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 64 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 65 * SUCH DAMAGE. 66 * 67 * @(#)kern_descrip.c 8.6 (Berkeley) 4/19/94 68 * $FreeBSD: src/sys/kern/kern_descrip.c,v 1.81.2.19 2004/02/28 00:43:31 tegge Exp $ 69 */ 70 71 #include <sys/param.h> 72 #include <sys/systm.h> 73 #include <sys/malloc.h> 74 #include <sys/sysproto.h> 75 #include <sys/conf.h> 76 #include <sys/device.h> 77 #include <sys/file.h> 78 #include <sys/filedesc.h> 79 #include <sys/kernel.h> 80 #include <sys/sysctl.h> 81 #include <sys/vnode.h> 82 #include <sys/proc.h> 83 #include <sys/nlookup.h> 84 #include <sys/stat.h> 85 #include <sys/filio.h> 86 #include <sys/fcntl.h> 87 #include <sys/unistd.h> 88 #include <sys/resourcevar.h> 89 #include <sys/event.h> 90 #include <sys/kern_syscall.h> 91 #include <sys/kcore.h> 92 #include <sys/kinfo.h> 93 #include <sys/un.h> 94 #include <sys/objcache.h> 95 96 #include <vm/vm.h> 97 #include <vm/vm_extern.h> 98 99 #include <sys/thread2.h> 100 #include <sys/file2.h> 101 #include <sys/spinlock2.h> 102 103 static int fdalloc_locked(struct proc *p, struct filedesc *fdp, 104 int want, int *result); 105 static void fsetfd_locked(struct filedesc *fdp, struct file *fp, int fd); 106 static void fdreserve_locked (struct filedesc *fdp, int fd0, int incr); 107 static struct file *funsetfd_locked (struct filedesc *fdp, int fd); 108 static void ffree(struct file *fp); 109 110 static MALLOC_DEFINE(M_FILEDESC, "file desc", "Open file descriptor table"); 111 static MALLOC_DEFINE(M_FILEDESC_TO_LEADER, "file desc to leader", 112 "file desc to leader structures"); 113 MALLOC_DEFINE(M_FILE, "file", "Open file structure"); 114 static MALLOC_DEFINE(M_SIGIO, "sigio", "sigio structures"); 115 116 static struct krate krate_uidinfo = { .freq = 1 }; 117 118 static d_open_t fdopen; 119 #define NUMFDESC 64 120 121 #define CDEV_MAJOR 22 122 static struct dev_ops fildesc_ops = { 123 { "FD", 0, 0 }, 124 .d_open = fdopen, 125 }; 126 127 /* 128 * Descriptor management. 129 */ 130 #ifndef NFILELIST_HEADS 131 #define NFILELIST_HEADS 257 /* primary number */ 132 #endif 133 134 struct filelist_head { 135 struct spinlock spin; 136 struct filelist list; 137 } __cachealign; 138 139 static struct filelist_head filelist_heads[NFILELIST_HEADS]; 140 141 static int nfiles; /* actual number of open files */ 142 extern int cmask; 143 144 struct lwkt_token revoke_token = LWKT_TOKEN_INITIALIZER(revoke_token); 145 146 static struct objcache *file_objcache; 147 148 static struct objcache_malloc_args file_malloc_args = { 149 .objsize = sizeof(struct file), 150 .mtype = M_FILE 151 }; 152 153 /* 154 * Fixup fd_freefile and fd_lastfile after a descriptor has been cleared. 155 * 156 * must be called with fdp->fd_spin exclusively held 157 */ 158 static __inline 159 void 160 fdfixup_locked(struct filedesc *fdp, int fd) 161 { 162 if (fd < fdp->fd_freefile) { 163 fdp->fd_freefile = fd; 164 } 165 while (fdp->fd_lastfile >= 0 && 166 fdp->fd_files[fdp->fd_lastfile].fp == NULL && 167 fdp->fd_files[fdp->fd_lastfile].reserved == 0 168 ) { 169 --fdp->fd_lastfile; 170 } 171 } 172 173 /* 174 * Clear the fd thread caches for this fdnode. 175 * 176 * If match_fdc is NULL, all thread caches of fdn will be cleared. 177 * The caller must hold fdp->fd_spin exclusively. The threads caching 178 * the descriptor do not have to be the current thread. The (status) 179 * argument is ignored. 180 * 181 * If match_fdc is not NULL, only the match_fdc's cache will be cleared. 182 * The caller must hold fdp->fd_spin shared and match_fdc must match a 183 * fdcache entry in curthread. match_fdc has been locked by the caller 184 * and had the specified (status). 185 * 186 * Since we are matching against a fp in the fdp (which must still be present 187 * at this time), fp will have at least two refs on any match and we can 188 * decrement the count trivially. 189 */ 190 static 191 void 192 fclearcache(struct fdnode *fdn, struct fdcache *match_fdc, int status) 193 { 194 struct fdcache *fdc; 195 struct file *fp; 196 int i; 197 198 /* 199 * match_fdc == NULL We are cleaning out all tdcache entries 200 * for the fdn and hold fdp->fd_spin exclusively. 201 * This can race against the target threads 202 * cleaning out specific entries. 203 * 204 * match_fdc != NULL We are cleaning out a specific tdcache 205 * entry on behalf of the owning thread 206 * and hold fdp->fd_spin shared. The thread 207 * has already locked the entry. This cannot 208 * race. 209 */ 210 fp = fdn->fp; 211 for (i = 0; i < NTDCACHEFD; ++i) { 212 if ((fdc = fdn->tdcache[i]) == NULL) 213 continue; 214 215 /* 216 * If match_fdc is non-NULL we are being asked to 217 * clear a specific fdc owned by curthread. There must 218 * be exactly one match. The caller has already locked 219 * the cache entry and will dispose of the lock after 220 * we return. 221 * 222 * Since we also have a shared lock on fdp, we 223 * can do this without atomic ops. 224 */ 225 if (match_fdc) { 226 if (fdc != match_fdc) 227 continue; 228 fdn->tdcache[i] = NULL; 229 KASSERT(fp == fdc->fp, 230 ("fclearcache(1): fp mismatch %p/%p\n", 231 fp, fdc->fp)); 232 fdc->fp = NULL; 233 fdc->fd = -1; 234 235 /* 236 * status can be 0 or 2. If 2 the ref is borrowed, 237 * if 0 the ref is not borrowed and we have to drop 238 * it. 239 */ 240 if (status == 0) 241 atomic_add_int(&fp->f_count, -1); 242 fdn->isfull = 0; /* heuristic */ 243 return; 244 } 245 246 /* 247 * Otherwise we hold an exclusive spin-lock and can only 248 * race thread consumers borrowing cache entries. 249 * 250 * Acquire the lock and dispose of the entry. We have to 251 * spin until we get the lock. 252 */ 253 for (;;) { 254 status = atomic_swap_int(&fdc->locked, 1); 255 if (status == 1) { /* foreign lock, retry */ 256 cpu_pause(); 257 continue; 258 } 259 fdn->tdcache[i] = NULL; 260 KASSERT(fp == fdc->fp, 261 ("fclearcache(2): fp mismatch %p/%p\n", 262 fp, fdc->fp)); 263 fdc->fp = NULL; 264 fdc->fd = -1; 265 if (status == 0) 266 atomic_add_int(&fp->f_count, -1); 267 fdn->isfull = 0; /* heuristic */ 268 atomic_swap_int(&fdc->locked, 0); 269 break; 270 } 271 } 272 KKASSERT(match_fdc == NULL); 273 } 274 275 /* 276 * Retrieve the fp for the specified fd given the specified file descriptor 277 * table. The fdp does not have to be owned by the current process. 278 * If flags != -1, fp->f_flag must contain at least one of the flags. 279 * 280 * This function is not able to cache the fp. 281 */ 282 struct file * 283 holdfp_fdp(struct filedesc *fdp, int fd, int flag) 284 { 285 struct file *fp; 286 287 spin_lock_shared(&fdp->fd_spin); 288 if (((u_int)fd) < fdp->fd_nfiles) { 289 fp = fdp->fd_files[fd].fp; /* can be NULL */ 290 if (fp) { 291 if ((fp->f_flag & flag) == 0 && flag != -1) { 292 fp = NULL; 293 } else { 294 fhold(fp); 295 } 296 } 297 } else { 298 fp = NULL; 299 } 300 spin_unlock_shared(&fdp->fd_spin); 301 302 return fp; 303 } 304 305 struct file * 306 holdfp_fdp_locked(struct filedesc *fdp, int fd, int flag) 307 { 308 struct file *fp; 309 310 if (((u_int)fd) < fdp->fd_nfiles) { 311 fp = fdp->fd_files[fd].fp; /* can be NULL */ 312 if (fp) { 313 if ((fp->f_flag & flag) == 0 && flag != -1) { 314 fp = NULL; 315 } else { 316 fhold(fp); 317 } 318 } 319 } else { 320 fp = NULL; 321 } 322 return fp; 323 } 324 325 /* 326 * Acquire the fp for the specified file descriptor, using the thread 327 * cache if possible and caching it if possible. 328 * 329 * td must be the curren thread. 330 */ 331 static 332 struct file * 333 _holdfp_cache(thread_t td, int fd) 334 { 335 struct filedesc *fdp; 336 struct fdcache *fdc; 337 struct fdcache *best; 338 struct fdnode *fdn; 339 struct file *fp; 340 int status; 341 int delta; 342 int i; 343 344 /* 345 * Fast 346 */ 347 for (fdc = &td->td_fdcache[0]; fdc < &td->td_fdcache[NFDCACHE]; ++fdc) { 348 if (fdc->fd != fd || fdc->fp == NULL) 349 continue; 350 status = atomic_swap_int(&fdc->locked, 1); 351 352 /* 353 * If someone else has locked our cache entry they are in 354 * the middle of clearing it, skip the entry. 355 */ 356 if (status == 1) 357 continue; 358 359 /* 360 * We have locked the entry, but if it no longer matches 361 * restore the previous state (0 or 2) and skip the entry. 362 */ 363 if (fdc->fd != fd || fdc->fp == NULL) { 364 atomic_swap_int(&fdc->locked, status); 365 continue; 366 } 367 368 /* 369 * We have locked a valid entry. We can borrow the ref 370 * for a mode 0 entry. We can get a valid fp for a mode 371 * 2 entry but not borrow the ref. 372 */ 373 if (status == 0) { 374 fp = fdc->fp; 375 fdc->lru = ++td->td_fdcache_lru; 376 atomic_swap_int(&fdc->locked, 2); 377 378 return fp; 379 } 380 if (status == 2) { 381 fp = fdc->fp; 382 fhold(fp); 383 fdc->lru = ++td->td_fdcache_lru; 384 atomic_swap_int(&fdc->locked, 2); 385 386 return fp; 387 } 388 KKASSERT(0); 389 } 390 391 /* 392 * Lookup the descriptor the slow way. This can contend against 393 * modifying operations in a multi-threaded environment and cause 394 * cache line ping ponging otherwise. 395 */ 396 fdp = td->td_proc->p_fd; 397 spin_lock_shared(&fdp->fd_spin); 398 399 if (((u_int)fd) < fdp->fd_nfiles) { 400 fp = fdp->fd_files[fd].fp; /* can be NULL */ 401 if (fp) { 402 fhold(fp); 403 if (fdp->fd_files[fd].isfull == 0) 404 goto enter; 405 } 406 } else { 407 fp = NULL; 408 } 409 spin_unlock_shared(&fdp->fd_spin); 410 411 return fp; 412 413 /* 414 * We found a valid fp and held it, fdp is still shared locked. 415 * Enter the fp into the per-thread cache. Find the oldest entry 416 * via lru, or an empty entry. 417 * 418 * Because fdp's spinlock is held (shared is fine), no other 419 * thread should be in the middle of clearing our selected entry. 420 */ 421 enter: 422 best = &td->td_fdcache[0]; 423 for (fdc = &td->td_fdcache[0]; fdc < &td->td_fdcache[NFDCACHE]; ++fdc) { 424 if (fdc->fp == NULL) { 425 best = fdc; 426 break; 427 } 428 delta = fdc->lru - best->lru; 429 if (delta < 0) 430 best = fdc; 431 } 432 433 /* 434 * Replace best 435 * 436 * Don't enter into the cache if we cannot get the lock. 437 */ 438 status = atomic_swap_int(&best->locked, 1); 439 if (status == 1) 440 goto done; 441 442 /* 443 * Clear the previous cache entry if present 444 */ 445 if (best->fp) { 446 KKASSERT(best->fd >= 0); 447 fclearcache(&fdp->fd_files[best->fd], best, status); 448 } 449 450 /* 451 * Create our new cache entry. This entry is 'safe' until we tie 452 * into the fdnode. If we cannot tie in, we will clear the entry. 453 */ 454 best->fd = fd; 455 best->fp = fp; 456 best->lru = ++td->td_fdcache_lru; 457 best->locked = 2; /* borrowed ref */ 458 459 fdn = &fdp->fd_files[fd]; 460 for (i = 0; i < NTDCACHEFD; ++i) { 461 if (fdn->tdcache[i] == NULL && 462 atomic_cmpset_ptr((void **)&fdn->tdcache[i], NULL, best)) { 463 goto done; 464 } 465 } 466 fdn->isfull = 1; /* no space */ 467 best->fd = -1; 468 best->fp = NULL; 469 best->locked = 0; 470 done: 471 spin_unlock_shared(&fdp->fd_spin); 472 473 return fp; 474 } 475 476 /* 477 * Drop the file pointer and return to the thread cache if possible. 478 * 479 * Caller must not hold fdp's spin lock. 480 * td must be the current thread. 481 */ 482 void 483 dropfp(thread_t td, int fd, struct file *fp) 484 { 485 struct filedesc *fdp; 486 struct fdcache *fdc; 487 int status; 488 489 fdp = td->td_proc->p_fd; 490 491 /* 492 * If our placeholder is still present we can re-cache the ref. 493 * 494 * Note that we can race an fclearcache(). 495 */ 496 for (fdc = &td->td_fdcache[0]; fdc < &td->td_fdcache[NFDCACHE]; ++fdc) { 497 if (fdc->fp != fp || fdc->fd != fd) 498 continue; 499 status = atomic_swap_int(&fdc->locked, 1); 500 switch(status) { 501 case 0: 502 /* 503 * Not in mode 2, fdrop fp without caching. 504 */ 505 atomic_swap_int(&fdc->locked, 0); 506 break; 507 case 1: 508 /* 509 * Not in mode 2, locked by someone else. 510 * fdrop fp without caching. 511 */ 512 break; 513 case 2: 514 /* 515 * Intact borrowed ref, return to mode 0 516 * indicating that we have returned the ref. 517 * 518 * Return the borrowed ref (2->1->0) 519 */ 520 if (fdc->fp == fp && fdc->fd == fd) { 521 atomic_swap_int(&fdc->locked, 0); 522 return; 523 } 524 atomic_swap_int(&fdc->locked, 2); 525 break; 526 } 527 } 528 529 /* 530 * Failed to re-cache, drop the fp without caching. 531 */ 532 fdrop(fp); 533 } 534 535 /* 536 * Clear all descriptors cached in the per-thread fd cache for 537 * the specified thread. 538 * 539 * Caller must not hold p_fd->spin. This function will temporarily 540 * obtain a shared spin lock. 541 */ 542 void 543 fexitcache(thread_t td) 544 { 545 struct filedesc *fdp; 546 struct fdcache *fdc; 547 int status; 548 int i; 549 550 if (td->td_proc == NULL) 551 return; 552 fdp = td->td_proc->p_fd; 553 if (fdp == NULL) 554 return; 555 556 /* 557 * A shared lock is sufficient as the caller controls td and we 558 * are only clearing td's cache. 559 */ 560 spin_lock_shared(&fdp->fd_spin); 561 for (i = 0; i < NFDCACHE; ++i) { 562 fdc = &td->td_fdcache[i]; 563 if (fdc->fp) { 564 status = atomic_swap_int(&fdc->locked, 1); 565 if (status == 1) { 566 cpu_pause(); 567 --i; 568 continue; 569 } 570 if (fdc->fp) { 571 KKASSERT(fdc->fd >= 0); 572 fclearcache(&fdp->fd_files[fdc->fd], fdc, 573 status); 574 } 575 atomic_swap_int(&fdc->locked, 0); 576 } 577 } 578 spin_unlock_shared(&fdp->fd_spin); 579 } 580 581 static __inline struct filelist_head * 582 fp2filelist(const struct file *fp) 583 { 584 u_int i; 585 586 i = (u_int)(uintptr_t)fp % NFILELIST_HEADS; 587 return &filelist_heads[i]; 588 } 589 590 static __inline 591 struct plimit * 592 readplimits(struct proc *p) 593 { 594 thread_t td = curthread; 595 struct plimit *limit; 596 597 limit = td->td_limit; 598 if (limit != p->p_limit) { 599 spin_lock_shared(&p->p_spin); 600 limit = p->p_limit; 601 atomic_add_int(&limit->p_refcnt, 1); 602 spin_unlock_shared(&p->p_spin); 603 if (td->td_limit) 604 plimit_free(td->td_limit); 605 td->td_limit = limit; 606 } 607 return limit; 608 } 609 610 /* 611 * System calls on descriptors. 612 */ 613 int 614 sys_getdtablesize(struct getdtablesize_args *uap) 615 { 616 struct proc *p = curproc; 617 struct plimit *limit = readplimits(p); 618 int dtsize; 619 620 if (limit->pl_rlimit[RLIMIT_NOFILE].rlim_cur > INT_MAX) 621 dtsize = INT_MAX; 622 else 623 dtsize = (int)limit->pl_rlimit[RLIMIT_NOFILE].rlim_cur; 624 625 if (dtsize > maxfilesperproc) 626 dtsize = maxfilesperproc; 627 if (dtsize < minfilesperproc) 628 dtsize = minfilesperproc; 629 if (p->p_ucred->cr_uid && dtsize > maxfilesperuser) 630 dtsize = maxfilesperuser; 631 uap->sysmsg_result = dtsize; 632 return (0); 633 } 634 635 /* 636 * Duplicate a file descriptor to a particular value. 637 * 638 * note: keep in mind that a potential race condition exists when closing 639 * descriptors from a shared descriptor table (via rfork). 640 */ 641 int 642 sys_dup2(struct dup2_args *uap) 643 { 644 int error; 645 int fd = 0; 646 647 error = kern_dup(DUP_FIXED, uap->from, uap->to, &fd); 648 uap->sysmsg_fds[0] = fd; 649 650 return (error); 651 } 652 653 /* 654 * Duplicate a file descriptor. 655 */ 656 int 657 sys_dup(struct dup_args *uap) 658 { 659 int error; 660 int fd = 0; 661 662 error = kern_dup(DUP_VARIABLE, uap->fd, 0, &fd); 663 uap->sysmsg_fds[0] = fd; 664 665 return (error); 666 } 667 668 /* 669 * MPALMOSTSAFE - acquires mplock for fp operations 670 */ 671 int 672 kern_fcntl(int fd, int cmd, union fcntl_dat *dat, struct ucred *cred) 673 { 674 struct thread *td = curthread; 675 struct proc *p = td->td_proc; 676 struct file *fp; 677 struct vnode *vp; 678 u_int newmin; 679 u_int oflags; 680 u_int nflags; 681 int closedcounter; 682 int tmp, error, flg = F_POSIX; 683 684 KKASSERT(p); 685 686 /* 687 * Operations on file descriptors that do not require a file pointer. 688 */ 689 switch (cmd) { 690 case F_GETFD: 691 error = fgetfdflags(p->p_fd, fd, &tmp); 692 if (error == 0) 693 dat->fc_cloexec = (tmp & UF_EXCLOSE) ? FD_CLOEXEC : 0; 694 return (error); 695 696 case F_SETFD: 697 if (dat->fc_cloexec & FD_CLOEXEC) 698 error = fsetfdflags(p->p_fd, fd, UF_EXCLOSE); 699 else 700 error = fclrfdflags(p->p_fd, fd, UF_EXCLOSE); 701 return (error); 702 case F_DUPFD: 703 newmin = dat->fc_fd; 704 error = kern_dup(DUP_VARIABLE | DUP_FCNTL, fd, newmin, 705 &dat->fc_fd); 706 return (error); 707 case F_DUPFD_CLOEXEC: 708 newmin = dat->fc_fd; 709 error = kern_dup(DUP_VARIABLE | DUP_CLOEXEC | DUP_FCNTL, 710 fd, newmin, &dat->fc_fd); 711 return (error); 712 case F_DUP2FD: 713 newmin = dat->fc_fd; 714 error = kern_dup(DUP_FIXED, fd, newmin, &dat->fc_fd); 715 return (error); 716 case F_DUP2FD_CLOEXEC: 717 newmin = dat->fc_fd; 718 error = kern_dup(DUP_FIXED | DUP_CLOEXEC, fd, newmin, 719 &dat->fc_fd); 720 return (error); 721 default: 722 break; 723 } 724 725 /* 726 * Operations on file pointers 727 */ 728 closedcounter = p->p_fd->fd_closedcounter; 729 if ((fp = holdfp(td, fd, -1)) == NULL) 730 return (EBADF); 731 732 switch (cmd) { 733 case F_GETFL: 734 dat->fc_flags = OFLAGS(fp->f_flag); 735 error = 0; 736 break; 737 738 case F_SETFL: 739 oflags = fp->f_flag; 740 nflags = FFLAGS(dat->fc_flags & ~O_ACCMODE) & FCNTLFLAGS; 741 nflags |= oflags & ~FCNTLFLAGS; 742 743 error = 0; 744 if (((nflags ^ oflags) & O_APPEND) && (oflags & FAPPENDONLY)) 745 error = EINVAL; 746 if (error == 0 && ((nflags ^ oflags) & FASYNC)) { 747 tmp = nflags & FASYNC; 748 error = fo_ioctl(fp, FIOASYNC, (caddr_t)&tmp, 749 cred, NULL); 750 } 751 752 /* 753 * If no error, must be atomically set. 754 */ 755 while (error == 0) { 756 oflags = fp->f_flag; 757 cpu_ccfence(); 758 nflags = (oflags & ~FCNTLFLAGS) | (nflags & FCNTLFLAGS); 759 if (atomic_cmpset_int(&fp->f_flag, oflags, nflags)) 760 break; 761 cpu_pause(); 762 } 763 break; 764 765 case F_GETOWN: 766 error = fo_ioctl(fp, FIOGETOWN, (caddr_t)&dat->fc_owner, 767 cred, NULL); 768 break; 769 770 case F_SETOWN: 771 error = fo_ioctl(fp, FIOSETOWN, (caddr_t)&dat->fc_owner, 772 cred, NULL); 773 break; 774 775 case F_SETLKW: 776 flg |= F_WAIT; 777 /* Fall into F_SETLK */ 778 779 case F_SETLK: 780 if (fp->f_type != DTYPE_VNODE) { 781 error = EBADF; 782 break; 783 } 784 vp = (struct vnode *)fp->f_data; 785 786 /* 787 * copyin/lockop may block 788 */ 789 if (dat->fc_flock.l_whence == SEEK_CUR) 790 dat->fc_flock.l_start += fp->f_offset; 791 792 switch (dat->fc_flock.l_type) { 793 case F_RDLCK: 794 if ((fp->f_flag & FREAD) == 0) { 795 error = EBADF; 796 break; 797 } 798 if (p->p_leader->p_advlock_flag == 0) 799 p->p_leader->p_advlock_flag = 1; 800 error = VOP_ADVLOCK(vp, (caddr_t)p->p_leader, F_SETLK, 801 &dat->fc_flock, flg); 802 break; 803 case F_WRLCK: 804 if ((fp->f_flag & FWRITE) == 0) { 805 error = EBADF; 806 break; 807 } 808 if (p->p_leader->p_advlock_flag == 0) 809 p->p_leader->p_advlock_flag = 1; 810 error = VOP_ADVLOCK(vp, (caddr_t)p->p_leader, F_SETLK, 811 &dat->fc_flock, flg); 812 break; 813 case F_UNLCK: 814 error = VOP_ADVLOCK(vp, (caddr_t)p->p_leader, F_UNLCK, 815 &dat->fc_flock, F_POSIX); 816 break; 817 default: 818 error = EINVAL; 819 break; 820 } 821 822 /* 823 * It is possible to race a close() on the descriptor while 824 * we were blocked getting the lock. If this occurs the 825 * close might not have caught the lock. 826 */ 827 if (checkfdclosed(td, p->p_fd, fd, fp, closedcounter)) { 828 dat->fc_flock.l_whence = SEEK_SET; 829 dat->fc_flock.l_start = 0; 830 dat->fc_flock.l_len = 0; 831 dat->fc_flock.l_type = F_UNLCK; 832 VOP_ADVLOCK(vp, (caddr_t)p->p_leader, 833 F_UNLCK, &dat->fc_flock, F_POSIX); 834 } 835 break; 836 837 case F_GETLK: 838 if (fp->f_type != DTYPE_VNODE) { 839 error = EBADF; 840 break; 841 } 842 vp = (struct vnode *)fp->f_data; 843 /* 844 * copyin/lockop may block 845 */ 846 if (dat->fc_flock.l_type != F_RDLCK && 847 dat->fc_flock.l_type != F_WRLCK && 848 dat->fc_flock.l_type != F_UNLCK) { 849 error = EINVAL; 850 break; 851 } 852 if (dat->fc_flock.l_whence == SEEK_CUR) 853 dat->fc_flock.l_start += fp->f_offset; 854 error = VOP_ADVLOCK(vp, (caddr_t)p->p_leader, F_GETLK, 855 &dat->fc_flock, F_POSIX); 856 break; 857 default: 858 error = EINVAL; 859 break; 860 } 861 862 fdrop(fp); 863 return (error); 864 } 865 866 /* 867 * The file control system call. 868 */ 869 int 870 sys_fcntl(struct fcntl_args *uap) 871 { 872 union fcntl_dat dat; 873 int error; 874 875 switch (uap->cmd) { 876 case F_DUPFD: 877 case F_DUP2FD: 878 case F_DUPFD_CLOEXEC: 879 case F_DUP2FD_CLOEXEC: 880 dat.fc_fd = uap->arg; 881 break; 882 case F_SETFD: 883 dat.fc_cloexec = uap->arg; 884 break; 885 case F_SETFL: 886 dat.fc_flags = uap->arg; 887 break; 888 case F_SETOWN: 889 dat.fc_owner = uap->arg; 890 break; 891 case F_SETLKW: 892 case F_SETLK: 893 case F_GETLK: 894 error = copyin((caddr_t)uap->arg, &dat.fc_flock, 895 sizeof(struct flock)); 896 if (error) 897 return (error); 898 break; 899 } 900 901 error = kern_fcntl(uap->fd, uap->cmd, &dat, curthread->td_ucred); 902 903 if (error == 0) { 904 switch (uap->cmd) { 905 case F_DUPFD: 906 case F_DUP2FD: 907 case F_DUPFD_CLOEXEC: 908 case F_DUP2FD_CLOEXEC: 909 uap->sysmsg_result = dat.fc_fd; 910 break; 911 case F_GETFD: 912 uap->sysmsg_result = dat.fc_cloexec; 913 break; 914 case F_GETFL: 915 uap->sysmsg_result = dat.fc_flags; 916 break; 917 case F_GETOWN: 918 uap->sysmsg_result = dat.fc_owner; 919 break; 920 case F_GETLK: 921 error = copyout(&dat.fc_flock, (caddr_t)uap->arg, 922 sizeof(struct flock)); 923 break; 924 } 925 } 926 927 return (error); 928 } 929 930 /* 931 * Common code for dup, dup2, and fcntl(F_DUPFD). 932 * 933 * There are four type flags: DUP_FCNTL, DUP_FIXED, DUP_VARIABLE, and 934 * DUP_CLOEXEC. 935 * 936 * DUP_FCNTL is for handling EINVAL vs. EBADF differences between 937 * fcntl()'s F_DUPFD and F_DUPFD_CLOEXEC and dup2() (per POSIX). 938 * The next two flags are mutually exclusive, and the fourth is optional. 939 * DUP_FIXED tells kern_dup() to destructively dup over an existing file 940 * descriptor if "new" is already open. DUP_VARIABLE tells kern_dup() 941 * to find the lowest unused file descriptor that is greater than or 942 * equal to "new". DUP_CLOEXEC, which works with either of the first 943 * two flags, sets the close-on-exec flag on the "new" file descriptor. 944 */ 945 int 946 kern_dup(int flags, int old, int new, int *res) 947 { 948 struct thread *td = curthread; 949 struct proc *p = td->td_proc; 950 struct plimit *limit = readplimits(p); 951 struct filedesc *fdp = p->p_fd; 952 struct file *fp; 953 struct file *delfp; 954 int oldflags; 955 int holdleaders; 956 int dtsize; 957 int error, newfd; 958 959 /* 960 * Verify that we have a valid descriptor to dup from and 961 * possibly to dup to. When the new descriptor is out of 962 * bounds, fcntl()'s F_DUPFD and F_DUPFD_CLOEXEC must 963 * return EINVAL, while dup2() returns EBADF in 964 * this case. 965 * 966 * NOTE: maxfilesperuser is not applicable to dup() 967 */ 968 retry: 969 if (limit->pl_rlimit[RLIMIT_NOFILE].rlim_cur > INT_MAX) 970 dtsize = INT_MAX; 971 else 972 dtsize = (int)limit->pl_rlimit[RLIMIT_NOFILE].rlim_cur; 973 if (dtsize > maxfilesperproc) 974 dtsize = maxfilesperproc; 975 if (dtsize < minfilesperproc) 976 dtsize = minfilesperproc; 977 978 if (new < 0 || new > dtsize) 979 return (flags & DUP_FCNTL ? EINVAL : EBADF); 980 981 spin_lock(&fdp->fd_spin); 982 if ((unsigned)old >= fdp->fd_nfiles || fdp->fd_files[old].fp == NULL) { 983 spin_unlock(&fdp->fd_spin); 984 return (EBADF); 985 } 986 if ((flags & DUP_FIXED) && old == new) { 987 *res = new; 988 if (flags & DUP_CLOEXEC) 989 fdp->fd_files[new].fileflags |= UF_EXCLOSE; 990 spin_unlock(&fdp->fd_spin); 991 return (0); 992 } 993 fp = fdp->fd_files[old].fp; 994 oldflags = fdp->fd_files[old].fileflags; 995 fhold(fp); 996 997 /* 998 * Allocate a new descriptor if DUP_VARIABLE, or expand the table 999 * if the requested descriptor is beyond the current table size. 1000 * 1001 * This can block. Retry if the source descriptor no longer matches 1002 * or if our expectation in the expansion case races. 1003 * 1004 * If we are not expanding or allocating a new decriptor, then reset 1005 * the target descriptor to a reserved state so we have a uniform 1006 * setup for the next code block. 1007 */ 1008 if ((flags & DUP_VARIABLE) || new >= fdp->fd_nfiles) { 1009 error = fdalloc_locked(p, fdp, new, &newfd); 1010 if (error) { 1011 spin_unlock(&fdp->fd_spin); 1012 fdrop(fp); 1013 return (error); 1014 } 1015 /* 1016 * Check for ripout 1017 */ 1018 if (old >= fdp->fd_nfiles || fdp->fd_files[old].fp != fp) { 1019 fsetfd_locked(fdp, NULL, newfd); 1020 spin_unlock(&fdp->fd_spin); 1021 fdrop(fp); 1022 goto retry; 1023 } 1024 /* 1025 * Check for expansion race 1026 */ 1027 if ((flags & DUP_VARIABLE) == 0 && new != newfd) { 1028 fsetfd_locked(fdp, NULL, newfd); 1029 spin_unlock(&fdp->fd_spin); 1030 fdrop(fp); 1031 goto retry; 1032 } 1033 /* 1034 * Check for ripout, newfd reused old (this case probably 1035 * can't occur). 1036 */ 1037 if (old == newfd) { 1038 fsetfd_locked(fdp, NULL, newfd); 1039 spin_unlock(&fdp->fd_spin); 1040 fdrop(fp); 1041 goto retry; 1042 } 1043 new = newfd; 1044 delfp = NULL; 1045 } else { 1046 if (fdp->fd_files[new].reserved) { 1047 spin_unlock(&fdp->fd_spin); 1048 fdrop(fp); 1049 kprintf("Warning: dup(): target descriptor %d is " 1050 "reserved, waiting for it to be resolved\n", 1051 new); 1052 tsleep(fdp, 0, "fdres", hz); 1053 goto retry; 1054 } 1055 1056 /* 1057 * If the target descriptor was never allocated we have 1058 * to allocate it. If it was we have to clean out the 1059 * old descriptor. delfp inherits the ref from the 1060 * descriptor table. 1061 */ 1062 ++fdp->fd_closedcounter; 1063 fclearcache(&fdp->fd_files[new], NULL, 0); 1064 ++fdp->fd_closedcounter; 1065 delfp = fdp->fd_files[new].fp; 1066 fdp->fd_files[new].fp = NULL; 1067 fdp->fd_files[new].reserved = 1; 1068 if (delfp == NULL) { 1069 fdreserve_locked(fdp, new, 1); 1070 if (new > fdp->fd_lastfile) 1071 fdp->fd_lastfile = new; 1072 } 1073 1074 } 1075 1076 /* 1077 * NOTE: still holding an exclusive spinlock 1078 */ 1079 1080 /* 1081 * If a descriptor is being overwritten we may hve to tell 1082 * fdfree() to sleep to ensure that all relevant process 1083 * leaders can be traversed in closef(). 1084 */ 1085 if (delfp != NULL && p->p_fdtol != NULL) { 1086 fdp->fd_holdleaderscount++; 1087 holdleaders = 1; 1088 } else { 1089 holdleaders = 0; 1090 } 1091 KASSERT(delfp == NULL || (flags & DUP_FIXED), 1092 ("dup() picked an open file")); 1093 1094 /* 1095 * Duplicate the source descriptor, update lastfile. If the new 1096 * descriptor was not allocated and we aren't replacing an existing 1097 * descriptor we have to mark the descriptor as being in use. 1098 * 1099 * The fd_files[] array inherits fp's hold reference. 1100 */ 1101 fsetfd_locked(fdp, fp, new); 1102 if ((flags & DUP_CLOEXEC) != 0) 1103 fdp->fd_files[new].fileflags = oldflags | UF_EXCLOSE; 1104 else 1105 fdp->fd_files[new].fileflags = oldflags & ~UF_EXCLOSE; 1106 spin_unlock(&fdp->fd_spin); 1107 fdrop(fp); 1108 *res = new; 1109 1110 /* 1111 * If we dup'd over a valid file, we now own the reference to it 1112 * and must dispose of it using closef() semantics (as if a 1113 * close() were performed on it). 1114 */ 1115 if (delfp) { 1116 if (SLIST_FIRST(&delfp->f_klist)) 1117 knote_fdclose(delfp, fdp, new); 1118 closef(delfp, p); 1119 if (holdleaders) { 1120 spin_lock(&fdp->fd_spin); 1121 fdp->fd_holdleaderscount--; 1122 if (fdp->fd_holdleaderscount == 0 && 1123 fdp->fd_holdleaderswakeup != 0) { 1124 fdp->fd_holdleaderswakeup = 0; 1125 spin_unlock(&fdp->fd_spin); 1126 wakeup(&fdp->fd_holdleaderscount); 1127 } else { 1128 spin_unlock(&fdp->fd_spin); 1129 } 1130 } 1131 } 1132 return (0); 1133 } 1134 1135 /* 1136 * If sigio is on the list associated with a process or process group, 1137 * disable signalling from the device, remove sigio from the list and 1138 * free sigio. 1139 */ 1140 void 1141 funsetown(struct sigio **sigiop) 1142 { 1143 struct pgrp *pgrp; 1144 struct proc *p; 1145 struct sigio *sigio; 1146 1147 if ((sigio = *sigiop) != NULL) { 1148 lwkt_gettoken(&sigio_token); /* protect sigio */ 1149 KKASSERT(sigiop == sigio->sio_myref); 1150 sigio = *sigiop; 1151 *sigiop = NULL; 1152 lwkt_reltoken(&sigio_token); 1153 } 1154 if (sigio == NULL) 1155 return; 1156 1157 if (sigio->sio_pgid < 0) { 1158 pgrp = sigio->sio_pgrp; 1159 sigio->sio_pgrp = NULL; 1160 lwkt_gettoken(&pgrp->pg_token); 1161 SLIST_REMOVE(&pgrp->pg_sigiolst, sigio, sigio, sio_pgsigio); 1162 lwkt_reltoken(&pgrp->pg_token); 1163 pgrel(pgrp); 1164 } else /* if ((*sigiop)->sio_pgid > 0) */ { 1165 p = sigio->sio_proc; 1166 sigio->sio_proc = NULL; 1167 PHOLD(p); 1168 lwkt_gettoken(&p->p_token); 1169 SLIST_REMOVE(&p->p_sigiolst, sigio, sigio, sio_pgsigio); 1170 lwkt_reltoken(&p->p_token); 1171 PRELE(p); 1172 } 1173 crfree(sigio->sio_ucred); 1174 sigio->sio_ucred = NULL; 1175 kfree(sigio, M_SIGIO); 1176 } 1177 1178 /* 1179 * Free a list of sigio structures. Caller is responsible for ensuring 1180 * that the list is MPSAFE. 1181 */ 1182 void 1183 funsetownlst(struct sigiolst *sigiolst) 1184 { 1185 struct sigio *sigio; 1186 1187 while ((sigio = SLIST_FIRST(sigiolst)) != NULL) 1188 funsetown(sigio->sio_myref); 1189 } 1190 1191 /* 1192 * This is common code for FIOSETOWN ioctl called by fcntl(fd, F_SETOWN, arg). 1193 * 1194 * After permission checking, add a sigio structure to the sigio list for 1195 * the process or process group. 1196 */ 1197 int 1198 fsetown(pid_t pgid, struct sigio **sigiop) 1199 { 1200 struct proc *proc = NULL; 1201 struct pgrp *pgrp = NULL; 1202 struct sigio *sigio; 1203 int error; 1204 1205 if (pgid == 0) { 1206 funsetown(sigiop); 1207 return (0); 1208 } 1209 1210 if (pgid > 0) { 1211 proc = pfind(pgid); 1212 if (proc == NULL) { 1213 error = ESRCH; 1214 goto done; 1215 } 1216 1217 /* 1218 * Policy - Don't allow a process to FSETOWN a process 1219 * in another session. 1220 * 1221 * Remove this test to allow maximum flexibility or 1222 * restrict FSETOWN to the current process or process 1223 * group for maximum safety. 1224 */ 1225 if (proc->p_session != curproc->p_session) { 1226 error = EPERM; 1227 goto done; 1228 } 1229 } else /* if (pgid < 0) */ { 1230 pgrp = pgfind(-pgid); 1231 if (pgrp == NULL) { 1232 error = ESRCH; 1233 goto done; 1234 } 1235 1236 /* 1237 * Policy - Don't allow a process to FSETOWN a process 1238 * in another session. 1239 * 1240 * Remove this test to allow maximum flexibility or 1241 * restrict FSETOWN to the current process or process 1242 * group for maximum safety. 1243 */ 1244 if (pgrp->pg_session != curproc->p_session) { 1245 error = EPERM; 1246 goto done; 1247 } 1248 } 1249 sigio = kmalloc(sizeof(struct sigio), M_SIGIO, M_WAITOK | M_ZERO); 1250 if (pgid > 0) { 1251 KKASSERT(pgrp == NULL); 1252 lwkt_gettoken(&proc->p_token); 1253 SLIST_INSERT_HEAD(&proc->p_sigiolst, sigio, sio_pgsigio); 1254 sigio->sio_proc = proc; 1255 lwkt_reltoken(&proc->p_token); 1256 } else { 1257 KKASSERT(proc == NULL); 1258 lwkt_gettoken(&pgrp->pg_token); 1259 SLIST_INSERT_HEAD(&pgrp->pg_sigiolst, sigio, sio_pgsigio); 1260 sigio->sio_pgrp = pgrp; 1261 lwkt_reltoken(&pgrp->pg_token); 1262 pgrp = NULL; 1263 } 1264 sigio->sio_pgid = pgid; 1265 sigio->sio_ucred = crhold(curthread->td_ucred); 1266 /* It would be convenient if p_ruid was in ucred. */ 1267 sigio->sio_ruid = sigio->sio_ucred->cr_ruid; 1268 sigio->sio_myref = sigiop; 1269 1270 lwkt_gettoken(&sigio_token); 1271 while (*sigiop) 1272 funsetown(sigiop); 1273 *sigiop = sigio; 1274 lwkt_reltoken(&sigio_token); 1275 error = 0; 1276 done: 1277 if (pgrp) 1278 pgrel(pgrp); 1279 if (proc) 1280 PRELE(proc); 1281 return (error); 1282 } 1283 1284 /* 1285 * This is common code for FIOGETOWN ioctl called by fcntl(fd, F_GETOWN, arg). 1286 */ 1287 pid_t 1288 fgetown(struct sigio **sigiop) 1289 { 1290 struct sigio *sigio; 1291 pid_t own; 1292 1293 lwkt_gettoken_shared(&sigio_token); 1294 sigio = *sigiop; 1295 own = (sigio != NULL ? sigio->sio_pgid : 0); 1296 lwkt_reltoken(&sigio_token); 1297 1298 return (own); 1299 } 1300 1301 /* 1302 * Close many file descriptors. 1303 */ 1304 int 1305 sys_closefrom(struct closefrom_args *uap) 1306 { 1307 return(kern_closefrom(uap->fd)); 1308 } 1309 1310 /* 1311 * Close all file descriptors greater then or equal to fd 1312 */ 1313 int 1314 kern_closefrom(int fd) 1315 { 1316 struct thread *td = curthread; 1317 struct proc *p = td->td_proc; 1318 struct filedesc *fdp; 1319 int error; 1320 int e2; 1321 1322 KKASSERT(p); 1323 fdp = p->p_fd; 1324 1325 if (fd < 0) 1326 return (EINVAL); 1327 1328 /* 1329 * NOTE: This function will skip unassociated descriptors and 1330 * reserved descriptors that have not yet been assigned. 1331 * fd_lastfile can change as a side effect of kern_close(). 1332 * 1333 * NOTE: We accumulate EINTR errors and return EINTR if any 1334 * close() returned EINTR. However, the descriptor is 1335 * still closed and we do not break out of the loop. 1336 */ 1337 error = 0; 1338 spin_lock(&fdp->fd_spin); 1339 while (fd <= fdp->fd_lastfile) { 1340 if (fdp->fd_files[fd].fp != NULL) { 1341 spin_unlock(&fdp->fd_spin); 1342 /* ok if this races another close */ 1343 e2 = kern_close(fd); 1344 if (e2 == EINTR) 1345 error = EINTR; 1346 spin_lock(&fdp->fd_spin); 1347 } 1348 ++fd; 1349 } 1350 spin_unlock(&fdp->fd_spin); 1351 1352 return error; 1353 } 1354 1355 /* 1356 * Close a file descriptor. 1357 */ 1358 int 1359 sys_close(struct close_args *uap) 1360 { 1361 return(kern_close(uap->fd)); 1362 } 1363 1364 /* 1365 * close() helper 1366 */ 1367 int 1368 kern_close(int fd) 1369 { 1370 struct thread *td = curthread; 1371 struct proc *p = td->td_proc; 1372 struct filedesc *fdp; 1373 struct file *fp; 1374 int error; 1375 int holdleaders; 1376 1377 KKASSERT(p); 1378 fdp = p->p_fd; 1379 1380 /* 1381 * funsetfd*() also clears the fd cache 1382 */ 1383 spin_lock(&fdp->fd_spin); 1384 if ((fp = funsetfd_locked(fdp, fd)) == NULL) { 1385 spin_unlock(&fdp->fd_spin); 1386 return (EBADF); 1387 } 1388 holdleaders = 0; 1389 if (p->p_fdtol != NULL) { 1390 /* 1391 * Ask fdfree() to sleep to ensure that all relevant 1392 * process leaders can be traversed in closef(). 1393 */ 1394 fdp->fd_holdleaderscount++; 1395 holdleaders = 1; 1396 } 1397 1398 /* 1399 * we now hold the fp reference that used to be owned by the descriptor 1400 * array. 1401 */ 1402 spin_unlock(&fdp->fd_spin); 1403 if (SLIST_FIRST(&fp->f_klist)) 1404 knote_fdclose(fp, fdp, fd); 1405 error = closef(fp, p); 1406 if (holdleaders) { 1407 spin_lock(&fdp->fd_spin); 1408 fdp->fd_holdleaderscount--; 1409 if (fdp->fd_holdleaderscount == 0 && 1410 fdp->fd_holdleaderswakeup != 0) { 1411 fdp->fd_holdleaderswakeup = 0; 1412 spin_unlock(&fdp->fd_spin); 1413 wakeup(&fdp->fd_holdleaderscount); 1414 } else { 1415 spin_unlock(&fdp->fd_spin); 1416 } 1417 } 1418 return (error); 1419 } 1420 1421 /* 1422 * shutdown_args(int fd, int how) 1423 */ 1424 int 1425 kern_shutdown(int fd, int how) 1426 { 1427 struct thread *td = curthread; 1428 struct file *fp; 1429 int error; 1430 1431 if ((fp = holdfp(td, fd, -1)) == NULL) 1432 return (EBADF); 1433 error = fo_shutdown(fp, how); 1434 fdrop(fp); 1435 1436 return (error); 1437 } 1438 1439 /* 1440 * MPALMOSTSAFE 1441 */ 1442 int 1443 sys_shutdown(struct shutdown_args *uap) 1444 { 1445 int error; 1446 1447 error = kern_shutdown(uap->s, uap->how); 1448 1449 return (error); 1450 } 1451 1452 /* 1453 * fstat() helper 1454 */ 1455 int 1456 kern_fstat(int fd, struct stat *ub) 1457 { 1458 struct thread *td = curthread; 1459 struct file *fp; 1460 int error; 1461 1462 if ((fp = holdfp(td, fd, -1)) == NULL) 1463 return (EBADF); 1464 error = fo_stat(fp, ub, td->td_ucred); 1465 fdrop(fp); 1466 1467 return (error); 1468 } 1469 1470 /* 1471 * Return status information about a file descriptor. 1472 */ 1473 int 1474 sys_fstat(struct fstat_args *uap) 1475 { 1476 struct stat st; 1477 int error; 1478 1479 error = kern_fstat(uap->fd, &st); 1480 1481 if (error == 0) 1482 error = copyout(&st, uap->sb, sizeof(st)); 1483 return (error); 1484 } 1485 1486 /* 1487 * Return pathconf information about a file descriptor. 1488 * 1489 * MPALMOSTSAFE 1490 */ 1491 int 1492 sys_fpathconf(struct fpathconf_args *uap) 1493 { 1494 struct thread *td = curthread; 1495 struct file *fp; 1496 struct vnode *vp; 1497 int error = 0; 1498 1499 if ((fp = holdfp(td, uap->fd, -1)) == NULL) 1500 return (EBADF); 1501 1502 switch (fp->f_type) { 1503 case DTYPE_PIPE: 1504 case DTYPE_SOCKET: 1505 if (uap->name != _PC_PIPE_BUF) { 1506 error = EINVAL; 1507 } else { 1508 uap->sysmsg_result = PIPE_BUF; 1509 error = 0; 1510 } 1511 break; 1512 case DTYPE_FIFO: 1513 case DTYPE_VNODE: 1514 vp = (struct vnode *)fp->f_data; 1515 error = VOP_PATHCONF(vp, uap->name, &uap->sysmsg_reg); 1516 break; 1517 default: 1518 error = EOPNOTSUPP; 1519 break; 1520 } 1521 fdrop(fp); 1522 return(error); 1523 } 1524 1525 /* 1526 * Grow the file table so it can hold through descriptor (want). 1527 * 1528 * The fdp's spinlock must be held exclusively on entry and may be held 1529 * exclusively on return. The spinlock may be cycled by the routine. 1530 */ 1531 static void 1532 fdgrow_locked(struct filedesc *fdp, int want) 1533 { 1534 struct fdnode *newfiles; 1535 struct fdnode *oldfiles; 1536 int nf, extra; 1537 1538 nf = fdp->fd_nfiles; 1539 do { 1540 /* nf has to be of the form 2^n - 1 */ 1541 nf = 2 * nf + 1; 1542 } while (nf <= want); 1543 1544 spin_unlock(&fdp->fd_spin); 1545 newfiles = kmalloc(nf * sizeof(struct fdnode), M_FILEDESC, M_WAITOK); 1546 spin_lock(&fdp->fd_spin); 1547 1548 /* 1549 * We could have raced another extend while we were not holding 1550 * the spinlock. 1551 */ 1552 if (fdp->fd_nfiles >= nf) { 1553 spin_unlock(&fdp->fd_spin); 1554 kfree(newfiles, M_FILEDESC); 1555 spin_lock(&fdp->fd_spin); 1556 return; 1557 } 1558 /* 1559 * Copy the existing ofile and ofileflags arrays 1560 * and zero the new portion of each array. 1561 */ 1562 extra = nf - fdp->fd_nfiles; 1563 bcopy(fdp->fd_files, newfiles, fdp->fd_nfiles * sizeof(struct fdnode)); 1564 bzero(&newfiles[fdp->fd_nfiles], extra * sizeof(struct fdnode)); 1565 1566 oldfiles = fdp->fd_files; 1567 fdp->fd_files = newfiles; 1568 fdp->fd_nfiles = nf; 1569 1570 if (oldfiles != fdp->fd_builtin_files) { 1571 spin_unlock(&fdp->fd_spin); 1572 kfree(oldfiles, M_FILEDESC); 1573 spin_lock(&fdp->fd_spin); 1574 } 1575 } 1576 1577 /* 1578 * Number of nodes in right subtree, including the root. 1579 */ 1580 static __inline int 1581 right_subtree_size(int n) 1582 { 1583 return (n ^ (n | (n + 1))); 1584 } 1585 1586 /* 1587 * Bigger ancestor. 1588 */ 1589 static __inline int 1590 right_ancestor(int n) 1591 { 1592 return (n | (n + 1)); 1593 } 1594 1595 /* 1596 * Smaller ancestor. 1597 */ 1598 static __inline int 1599 left_ancestor(int n) 1600 { 1601 return ((n & (n + 1)) - 1); 1602 } 1603 1604 /* 1605 * Traverse the in-place binary tree buttom-up adjusting the allocation 1606 * count so scans can determine where free descriptors are located. 1607 * 1608 * caller must be holding an exclusive spinlock on fdp 1609 */ 1610 static 1611 void 1612 fdreserve_locked(struct filedesc *fdp, int fd, int incr) 1613 { 1614 while (fd >= 0) { 1615 fdp->fd_files[fd].allocated += incr; 1616 KKASSERT(fdp->fd_files[fd].allocated >= 0); 1617 fd = left_ancestor(fd); 1618 } 1619 } 1620 1621 /* 1622 * Reserve a file descriptor for the process. If no error occurs, the 1623 * caller MUST at some point call fsetfd() or assign a file pointer 1624 * or dispose of the reservation. 1625 */ 1626 static 1627 int 1628 fdalloc_locked(struct proc *p, struct filedesc *fdp, int want, int *result) 1629 { 1630 struct plimit *limit = readplimits(p); 1631 struct uidinfo *uip; 1632 int fd, rsize, rsum, node, lim; 1633 1634 /* 1635 * Check dtable size limit 1636 */ 1637 *result = -1; /* avoid gcc warnings */ 1638 if (limit->pl_rlimit[RLIMIT_NOFILE].rlim_cur > INT_MAX) 1639 lim = INT_MAX; 1640 else 1641 lim = (int)limit->pl_rlimit[RLIMIT_NOFILE].rlim_cur; 1642 1643 if (lim > maxfilesperproc) 1644 lim = maxfilesperproc; 1645 if (lim < minfilesperproc) 1646 lim = minfilesperproc; 1647 if (want >= lim) 1648 return (EMFILE); 1649 1650 /* 1651 * Check that the user has not run out of descriptors (non-root only). 1652 * As a safety measure the dtable is allowed to have at least 1653 * minfilesperproc open fds regardless of the maxfilesperuser limit. 1654 * 1655 * This isn't as loose a spec as ui_posixlocks, so we use atomic 1656 * ops to force synchronize and recheck if we would otherwise 1657 * error. 1658 */ 1659 if (p->p_ucred->cr_uid && fdp->fd_nfiles >= minfilesperproc) { 1660 uip = p->p_ucred->cr_uidinfo; 1661 if (uip->ui_openfiles > maxfilesperuser) { 1662 int n; 1663 int count; 1664 1665 count = 0; 1666 for (n = 0; n < ncpus; ++n) { 1667 count += atomic_swap_int( 1668 &uip->ui_pcpu[n].pu_openfiles, 0); 1669 } 1670 atomic_add_int(&uip->ui_openfiles, count); 1671 if (uip->ui_openfiles > maxfilesperuser) { 1672 krateprintf(&krate_uidinfo, 1673 "Warning: user %d pid %d (%s) " 1674 "ran out of file descriptors " 1675 "(%d/%d)\n", 1676 p->p_ucred->cr_uid, (int)p->p_pid, 1677 p->p_comm, 1678 uip->ui_openfiles, maxfilesperuser); 1679 return(ENFILE); 1680 } 1681 } 1682 } 1683 1684 /* 1685 * Grow the dtable if necessary 1686 */ 1687 if (want >= fdp->fd_nfiles) 1688 fdgrow_locked(fdp, want); 1689 1690 /* 1691 * Search for a free descriptor starting at the higher 1692 * of want or fd_freefile. If that fails, consider 1693 * expanding the ofile array. 1694 * 1695 * NOTE! the 'allocated' field is a cumulative recursive allocation 1696 * count. If we happen to see a value of 0 then we can shortcut 1697 * our search. Otherwise we run through through the tree going 1698 * down branches we know have free descriptor(s) until we hit a 1699 * leaf node. The leaf node will be free but will not necessarily 1700 * have an allocated field of 0. 1701 */ 1702 retry: 1703 /* move up the tree looking for a subtree with a free node */ 1704 for (fd = max(want, fdp->fd_freefile); fd < min(fdp->fd_nfiles, lim); 1705 fd = right_ancestor(fd)) { 1706 if (fdp->fd_files[fd].allocated == 0) 1707 goto found; 1708 1709 rsize = right_subtree_size(fd); 1710 if (fdp->fd_files[fd].allocated == rsize) 1711 continue; /* right subtree full */ 1712 1713 /* 1714 * Free fd is in the right subtree of the tree rooted at fd. 1715 * Call that subtree R. Look for the smallest (leftmost) 1716 * subtree of R with an unallocated fd: continue moving 1717 * down the left branch until encountering a full left 1718 * subtree, then move to the right. 1719 */ 1720 for (rsum = 0, rsize /= 2; rsize > 0; rsize /= 2) { 1721 node = fd + rsize; 1722 rsum += fdp->fd_files[node].allocated; 1723 if (fdp->fd_files[fd].allocated == rsum + rsize) { 1724 fd = node; /* move to the right */ 1725 if (fdp->fd_files[node].allocated == 0) 1726 goto found; 1727 rsum = 0; 1728 } 1729 } 1730 goto found; 1731 } 1732 1733 /* 1734 * No space in current array. Expand? 1735 */ 1736 if (fdp->fd_nfiles >= lim) { 1737 return (EMFILE); 1738 } 1739 fdgrow_locked(fdp, want); 1740 goto retry; 1741 1742 found: 1743 KKASSERT(fd < fdp->fd_nfiles); 1744 if (fd > fdp->fd_lastfile) 1745 fdp->fd_lastfile = fd; 1746 if (want <= fdp->fd_freefile) 1747 fdp->fd_freefile = fd; 1748 *result = fd; 1749 KKASSERT(fdp->fd_files[fd].fp == NULL); 1750 KKASSERT(fdp->fd_files[fd].reserved == 0); 1751 fdp->fd_files[fd].fileflags = 0; 1752 fdp->fd_files[fd].reserved = 1; 1753 fdreserve_locked(fdp, fd, 1); 1754 1755 return (0); 1756 } 1757 1758 int 1759 fdalloc(struct proc *p, int want, int *result) 1760 { 1761 struct filedesc *fdp = p->p_fd; 1762 int error; 1763 1764 spin_lock(&fdp->fd_spin); 1765 error = fdalloc_locked(p, fdp, want, result); 1766 spin_unlock(&fdp->fd_spin); 1767 1768 return error; 1769 } 1770 1771 /* 1772 * Check to see whether n user file descriptors 1773 * are available to the process p. 1774 */ 1775 int 1776 fdavail(struct proc *p, int n) 1777 { 1778 struct plimit *limit = readplimits(p); 1779 struct filedesc *fdp = p->p_fd; 1780 struct fdnode *fdnode; 1781 int i, lim, last; 1782 1783 if (limit->pl_rlimit[RLIMIT_NOFILE].rlim_cur > INT_MAX) 1784 lim = INT_MAX; 1785 else 1786 lim = (int)limit->pl_rlimit[RLIMIT_NOFILE].rlim_cur; 1787 1788 if (lim > maxfilesperproc) 1789 lim = maxfilesperproc; 1790 if (lim < minfilesperproc) 1791 lim = minfilesperproc; 1792 1793 spin_lock(&fdp->fd_spin); 1794 if ((i = lim - fdp->fd_nfiles) > 0 && (n -= i) <= 0) { 1795 spin_unlock(&fdp->fd_spin); 1796 return (1); 1797 } 1798 last = min(fdp->fd_nfiles, lim); 1799 fdnode = &fdp->fd_files[fdp->fd_freefile]; 1800 for (i = last - fdp->fd_freefile; --i >= 0; ++fdnode) { 1801 if (fdnode->fp == NULL && --n <= 0) { 1802 spin_unlock(&fdp->fd_spin); 1803 return (1); 1804 } 1805 } 1806 spin_unlock(&fdp->fd_spin); 1807 return (0); 1808 } 1809 1810 /* 1811 * Revoke open descriptors referencing (f_data, f_type) 1812 * 1813 * Any revoke executed within a prison is only able to 1814 * revoke descriptors for processes within that prison. 1815 * 1816 * Returns 0 on success or an error code. 1817 */ 1818 struct fdrevoke_info { 1819 void *data; 1820 short type; 1821 short unused; 1822 int found; 1823 struct ucred *cred; 1824 struct file *nfp; 1825 }; 1826 1827 static int fdrevoke_check_callback(struct file *fp, void *vinfo); 1828 static int fdrevoke_proc_callback(struct proc *p, void *vinfo); 1829 1830 int 1831 fdrevoke(void *f_data, short f_type, struct ucred *cred) 1832 { 1833 struct fdrevoke_info info; 1834 int error; 1835 1836 bzero(&info, sizeof(info)); 1837 info.data = f_data; 1838 info.type = f_type; 1839 info.cred = cred; 1840 error = falloc(NULL, &info.nfp, NULL); 1841 if (error) 1842 return (error); 1843 1844 /* 1845 * Scan the file pointer table once. dups do not dup file pointers, 1846 * only descriptors, so there is no leak. Set FREVOKED on the fps 1847 * being revoked. 1848 * 1849 * Any fps sent over unix-domain sockets will be revoked by the 1850 * socket code checking for FREVOKED when the fps are externialized. 1851 * revoke_token is used to make sure that fps marked FREVOKED and 1852 * externalized will be picked up by the following allproc_scan(). 1853 */ 1854 lwkt_gettoken(&revoke_token); 1855 allfiles_scan_exclusive(fdrevoke_check_callback, &info); 1856 lwkt_reltoken(&revoke_token); 1857 1858 /* 1859 * If any fps were marked track down the related descriptors 1860 * and close them. Any dup()s at this point will notice 1861 * the FREVOKED already set in the fp and do the right thing. 1862 */ 1863 if (info.found) 1864 allproc_scan(fdrevoke_proc_callback, &info, 0); 1865 fdrop(info.nfp); 1866 return(0); 1867 } 1868 1869 /* 1870 * Locate matching file pointers directly. 1871 * 1872 * WARNING: allfiles_scan_exclusive() holds a spinlock through these calls! 1873 */ 1874 static int 1875 fdrevoke_check_callback(struct file *fp, void *vinfo) 1876 { 1877 struct fdrevoke_info *info = vinfo; 1878 1879 /* 1880 * File pointers already flagged for revokation are skipped. 1881 */ 1882 if (fp->f_flag & FREVOKED) 1883 return(0); 1884 1885 /* 1886 * If revoking from a prison file pointers created outside of 1887 * that prison, or file pointers without creds, cannot be revoked. 1888 */ 1889 if (info->cred->cr_prison && 1890 (fp->f_cred == NULL || 1891 info->cred->cr_prison != fp->f_cred->cr_prison)) { 1892 return(0); 1893 } 1894 1895 /* 1896 * If the file pointer matches then mark it for revocation. The 1897 * flag is currently only used by unp_revoke_gc(). 1898 * 1899 * info->found is a heuristic and can race in a SMP environment. 1900 */ 1901 if (info->data == fp->f_data && info->type == fp->f_type) { 1902 atomic_set_int(&fp->f_flag, FREVOKED); 1903 info->found = 1; 1904 } 1905 return(0); 1906 } 1907 1908 /* 1909 * Locate matching file pointers via process descriptor tables. 1910 */ 1911 static int 1912 fdrevoke_proc_callback(struct proc *p, void *vinfo) 1913 { 1914 struct fdrevoke_info *info = vinfo; 1915 struct filedesc *fdp; 1916 struct file *fp; 1917 int n; 1918 1919 if (p->p_stat == SIDL || p->p_stat == SZOMB) 1920 return(0); 1921 if (info->cred->cr_prison && 1922 info->cred->cr_prison != p->p_ucred->cr_prison) { 1923 return(0); 1924 } 1925 1926 /* 1927 * If the controlling terminal of the process matches the 1928 * vnode being revoked we clear the controlling terminal. 1929 * 1930 * The normal spec_close() may not catch this because it 1931 * uses curproc instead of p. 1932 */ 1933 if (p->p_session && info->type == DTYPE_VNODE && 1934 info->data == p->p_session->s_ttyvp) { 1935 p->p_session->s_ttyvp = NULL; 1936 vrele(info->data); 1937 } 1938 1939 /* 1940 * Softref the fdp to prevent it from being destroyed 1941 */ 1942 spin_lock(&p->p_spin); 1943 if ((fdp = p->p_fd) == NULL) { 1944 spin_unlock(&p->p_spin); 1945 return(0); 1946 } 1947 atomic_add_int(&fdp->fd_softrefs, 1); 1948 spin_unlock(&p->p_spin); 1949 1950 /* 1951 * Locate and close any matching file descriptors, replacing 1952 * them with info->nfp. 1953 */ 1954 spin_lock(&fdp->fd_spin); 1955 for (n = 0; n < fdp->fd_nfiles; ++n) { 1956 if ((fp = fdp->fd_files[n].fp) == NULL) 1957 continue; 1958 if (fp->f_flag & FREVOKED) { 1959 ++fdp->fd_closedcounter; 1960 fclearcache(&fdp->fd_files[n], NULL, 0); 1961 ++fdp->fd_closedcounter; 1962 fhold(info->nfp); 1963 fdp->fd_files[n].fp = info->nfp; 1964 spin_unlock(&fdp->fd_spin); 1965 knote_fdclose(fp, fdp, n); /* XXX */ 1966 closef(fp, p); 1967 spin_lock(&fdp->fd_spin); 1968 } 1969 } 1970 spin_unlock(&fdp->fd_spin); 1971 atomic_subtract_int(&fdp->fd_softrefs, 1); 1972 return(0); 1973 } 1974 1975 /* 1976 * falloc: 1977 * Create a new open file structure and reserve a file decriptor 1978 * for the process that refers to it. 1979 * 1980 * Root creds are checked using lp, or assumed if lp is NULL. If 1981 * resultfd is non-NULL then lp must also be non-NULL. No file 1982 * descriptor is reserved (and no process context is needed) if 1983 * resultfd is NULL. 1984 * 1985 * A file pointer with a refcount of 1 is returned. Note that the 1986 * file pointer is NOT associated with the descriptor. If falloc 1987 * returns success, fsetfd() MUST be called to either associate the 1988 * file pointer or clear the reservation. 1989 */ 1990 int 1991 falloc(struct lwp *lp, struct file **resultfp, int *resultfd) 1992 { 1993 static struct timeval lastfail; 1994 static int curfail; 1995 struct filelist_head *head; 1996 struct file *fp; 1997 struct ucred *cred = lp ? lp->lwp_thread->td_ucred : proc0.p_ucred; 1998 int error; 1999 2000 fp = NULL; 2001 2002 /* 2003 * Handle filetable full issues and root overfill. 2004 */ 2005 if (nfiles >= maxfiles - maxfilesrootres && 2006 (cred->cr_ruid != 0 || nfiles >= maxfiles)) { 2007 if (ppsratecheck(&lastfail, &curfail, 1)) { 2008 kprintf("kern.maxfiles limit exceeded by uid %d, " 2009 "please see tuning(7).\n", 2010 cred->cr_ruid); 2011 } 2012 error = ENFILE; 2013 goto done; 2014 } 2015 2016 /* 2017 * Allocate a new file descriptor. 2018 */ 2019 fp = objcache_get(file_objcache, M_WAITOK); 2020 bzero(fp, sizeof(*fp)); 2021 spin_init(&fp->f_spin, "falloc"); 2022 SLIST_INIT(&fp->f_klist); 2023 fp->f_count = 1; 2024 fp->f_ops = &badfileops; 2025 fp->f_seqcount = 1; 2026 fsetcred(fp, cred); 2027 atomic_add_int(&nfiles, 1); 2028 2029 head = fp2filelist(fp); 2030 spin_lock(&head->spin); 2031 LIST_INSERT_HEAD(&head->list, fp, f_list); 2032 spin_unlock(&head->spin); 2033 2034 if (resultfd) { 2035 if ((error = fdalloc(lp->lwp_proc, 0, resultfd)) != 0) { 2036 fdrop(fp); 2037 fp = NULL; 2038 } 2039 } else { 2040 error = 0; 2041 } 2042 done: 2043 *resultfp = fp; 2044 return (error); 2045 } 2046 2047 /* 2048 * Check for races against a file descriptor by determining that the 2049 * file pointer is still associated with the specified file descriptor, 2050 * and a close is not currently in progress. 2051 */ 2052 int 2053 checkfdclosed(thread_t td, struct filedesc *fdp, int fd, struct file *fp, 2054 int closedcounter) 2055 { 2056 struct fdcache *fdc; 2057 int error; 2058 2059 cpu_lfence(); 2060 if (fdp->fd_closedcounter == closedcounter) 2061 return 0; 2062 2063 if (td->td_proc && td->td_proc->p_fd == fdp) { 2064 for (fdc = &td->td_fdcache[0]; 2065 fdc < &td->td_fdcache[NFDCACHE]; ++fdc) { 2066 if (fdc->fd == fd && fdc->fp == fp) 2067 return 0; 2068 } 2069 } 2070 2071 spin_lock_shared(&fdp->fd_spin); 2072 if ((unsigned)fd >= fdp->fd_nfiles || fp != fdp->fd_files[fd].fp) 2073 error = EBADF; 2074 else 2075 error = 0; 2076 spin_unlock_shared(&fdp->fd_spin); 2077 return (error); 2078 } 2079 2080 /* 2081 * Associate a file pointer with a previously reserved file descriptor. 2082 * This function always succeeds. 2083 * 2084 * If fp is NULL, the file descriptor is returned to the pool. 2085 * 2086 * Caller must hold an exclusive spinlock on fdp->fd_spin. 2087 */ 2088 static void 2089 fsetfd_locked(struct filedesc *fdp, struct file *fp, int fd) 2090 { 2091 KKASSERT((unsigned)fd < fdp->fd_nfiles); 2092 KKASSERT(fdp->fd_files[fd].reserved != 0); 2093 if (fp) { 2094 fhold(fp); 2095 /* fclearcache(&fdp->fd_files[fd], NULL, 0); */ 2096 fdp->fd_files[fd].fp = fp; 2097 fdp->fd_files[fd].reserved = 0; 2098 } else { 2099 fdp->fd_files[fd].reserved = 0; 2100 fdreserve_locked(fdp, fd, -1); 2101 fdfixup_locked(fdp, fd); 2102 } 2103 } 2104 2105 /* 2106 * Caller must hold an exclusive spinlock on fdp->fd_spin. 2107 */ 2108 void 2109 fsetfd(struct filedesc *fdp, struct file *fp, int fd) 2110 { 2111 spin_lock(&fdp->fd_spin); 2112 fsetfd_locked(fdp, fp, fd); 2113 spin_unlock(&fdp->fd_spin); 2114 } 2115 2116 /* 2117 * Caller must hold an exclusive spinlock on fdp->fd_spin. 2118 */ 2119 static 2120 struct file * 2121 funsetfd_locked(struct filedesc *fdp, int fd) 2122 { 2123 struct file *fp; 2124 2125 if ((unsigned)fd >= fdp->fd_nfiles) 2126 return (NULL); 2127 if ((fp = fdp->fd_files[fd].fp) == NULL) 2128 return (NULL); 2129 ++fdp->fd_closedcounter; 2130 fclearcache(&fdp->fd_files[fd], NULL, 0); 2131 fdp->fd_files[fd].fp = NULL; 2132 fdp->fd_files[fd].fileflags = 0; 2133 ++fdp->fd_closedcounter; 2134 2135 fdreserve_locked(fdp, fd, -1); 2136 fdfixup_locked(fdp, fd); 2137 2138 return(fp); 2139 } 2140 2141 /* 2142 * WARNING: May not be called before initial fsetfd(). 2143 */ 2144 int 2145 fgetfdflags(struct filedesc *fdp, int fd, int *flagsp) 2146 { 2147 int error; 2148 2149 spin_lock_shared(&fdp->fd_spin); 2150 if (((u_int)fd) >= fdp->fd_nfiles) { 2151 error = EBADF; 2152 } else if (fdp->fd_files[fd].fp == NULL) { 2153 error = EBADF; 2154 } else { 2155 *flagsp = fdp->fd_files[fd].fileflags; 2156 error = 0; 2157 } 2158 spin_unlock_shared(&fdp->fd_spin); 2159 2160 return (error); 2161 } 2162 2163 /* 2164 * WARNING: May not be called before initial fsetfd(). 2165 */ 2166 int 2167 fsetfdflags(struct filedesc *fdp, int fd, int add_flags) 2168 { 2169 int error; 2170 2171 spin_lock(&fdp->fd_spin); 2172 if (((u_int)fd) >= fdp->fd_nfiles) { 2173 error = EBADF; 2174 } else if (fdp->fd_files[fd].fp == NULL) { 2175 error = EBADF; 2176 } else { 2177 fdp->fd_files[fd].fileflags |= add_flags; 2178 error = 0; 2179 } 2180 spin_unlock(&fdp->fd_spin); 2181 2182 return (error); 2183 } 2184 2185 /* 2186 * WARNING: May not be called before initial fsetfd(). 2187 */ 2188 int 2189 fclrfdflags(struct filedesc *fdp, int fd, int rem_flags) 2190 { 2191 int error; 2192 2193 spin_lock(&fdp->fd_spin); 2194 if (((u_int)fd) >= fdp->fd_nfiles) { 2195 error = EBADF; 2196 } else if (fdp->fd_files[fd].fp == NULL) { 2197 error = EBADF; 2198 } else { 2199 fdp->fd_files[fd].fileflags &= ~rem_flags; 2200 error = 0; 2201 } 2202 spin_unlock(&fdp->fd_spin); 2203 2204 return (error); 2205 } 2206 2207 /* 2208 * Set/Change/Clear the creds for a fp and synchronize the uidinfo. 2209 */ 2210 void 2211 fsetcred(struct file *fp, struct ucred *ncr) 2212 { 2213 struct ucred *ocr; 2214 struct uidinfo *uip; 2215 struct uidcount *pup; 2216 int cpu = mycpuid; 2217 int count; 2218 2219 ocr = fp->f_cred; 2220 if (ocr == NULL || ncr == NULL || ocr->cr_uidinfo != ncr->cr_uidinfo) { 2221 if (ocr) { 2222 uip = ocr->cr_uidinfo; 2223 pup = &uip->ui_pcpu[cpu]; 2224 atomic_add_int(&pup->pu_openfiles, -1); 2225 if (pup->pu_openfiles < -PUP_LIMIT || 2226 pup->pu_openfiles > PUP_LIMIT) { 2227 count = atomic_swap_int(&pup->pu_openfiles, 0); 2228 atomic_add_int(&uip->ui_openfiles, count); 2229 } 2230 } 2231 if (ncr) { 2232 uip = ncr->cr_uidinfo; 2233 pup = &uip->ui_pcpu[cpu]; 2234 atomic_add_int(&pup->pu_openfiles, 1); 2235 if (pup->pu_openfiles < -PUP_LIMIT || 2236 pup->pu_openfiles > PUP_LIMIT) { 2237 count = atomic_swap_int(&pup->pu_openfiles, 0); 2238 atomic_add_int(&uip->ui_openfiles, count); 2239 } 2240 } 2241 } 2242 if (ncr) 2243 crhold(ncr); 2244 fp->f_cred = ncr; 2245 if (ocr) 2246 crfree(ocr); 2247 } 2248 2249 /* 2250 * Free a file descriptor. 2251 */ 2252 static 2253 void 2254 ffree(struct file *fp) 2255 { 2256 KASSERT((fp->f_count == 0), ("ffree: fp_fcount not 0!")); 2257 fsetcred(fp, NULL); 2258 if (fp->f_nchandle.ncp) 2259 cache_drop(&fp->f_nchandle); 2260 objcache_put(file_objcache, fp); 2261 } 2262 2263 /* 2264 * called from init_main, initialize filedesc0 for proc0. 2265 */ 2266 void 2267 fdinit_bootstrap(struct proc *p0, struct filedesc *fdp0, int cmask) 2268 { 2269 p0->p_fd = fdp0; 2270 p0->p_fdtol = NULL; 2271 fdp0->fd_refcnt = 1; 2272 fdp0->fd_cmask = cmask; 2273 fdp0->fd_files = fdp0->fd_builtin_files; 2274 fdp0->fd_nfiles = NDFILE; 2275 fdp0->fd_lastfile = -1; 2276 spin_init(&fdp0->fd_spin, "fdinitbootstrap"); 2277 } 2278 2279 /* 2280 * Build a new filedesc structure. 2281 */ 2282 struct filedesc * 2283 fdinit(struct proc *p) 2284 { 2285 struct filedesc *newfdp; 2286 struct filedesc *fdp = p->p_fd; 2287 2288 newfdp = kmalloc(sizeof(struct filedesc), M_FILEDESC, M_WAITOK|M_ZERO); 2289 spin_lock(&fdp->fd_spin); 2290 if (fdp->fd_cdir) { 2291 newfdp->fd_cdir = fdp->fd_cdir; 2292 vref(newfdp->fd_cdir); 2293 cache_copy(&fdp->fd_ncdir, &newfdp->fd_ncdir); 2294 } 2295 2296 /* 2297 * rdir may not be set in e.g. proc0 or anything vm_fork'd off of 2298 * proc0, but should unconditionally exist in other processes. 2299 */ 2300 if (fdp->fd_rdir) { 2301 newfdp->fd_rdir = fdp->fd_rdir; 2302 vref(newfdp->fd_rdir); 2303 cache_copy(&fdp->fd_nrdir, &newfdp->fd_nrdir); 2304 } 2305 if (fdp->fd_jdir) { 2306 newfdp->fd_jdir = fdp->fd_jdir; 2307 vref(newfdp->fd_jdir); 2308 cache_copy(&fdp->fd_njdir, &newfdp->fd_njdir); 2309 } 2310 spin_unlock(&fdp->fd_spin); 2311 2312 /* Create the file descriptor table. */ 2313 newfdp->fd_refcnt = 1; 2314 newfdp->fd_cmask = cmask; 2315 newfdp->fd_files = newfdp->fd_builtin_files; 2316 newfdp->fd_nfiles = NDFILE; 2317 newfdp->fd_lastfile = -1; 2318 spin_init(&newfdp->fd_spin, "fdinit"); 2319 2320 return (newfdp); 2321 } 2322 2323 /* 2324 * Share a filedesc structure. 2325 */ 2326 struct filedesc * 2327 fdshare(struct proc *p) 2328 { 2329 struct filedesc *fdp; 2330 2331 fdp = p->p_fd; 2332 spin_lock(&fdp->fd_spin); 2333 fdp->fd_refcnt++; 2334 spin_unlock(&fdp->fd_spin); 2335 return (fdp); 2336 } 2337 2338 /* 2339 * Copy a filedesc structure. 2340 */ 2341 int 2342 fdcopy(struct proc *p, struct filedesc **fpp) 2343 { 2344 struct filedesc *fdp = p->p_fd; 2345 struct filedesc *newfdp; 2346 struct fdnode *fdnode; 2347 int i; 2348 int ni; 2349 2350 /* 2351 * Certain daemons might not have file descriptors. 2352 */ 2353 if (fdp == NULL) 2354 return (0); 2355 2356 /* 2357 * Allocate the new filedesc and fd_files[] array. This can race 2358 * with operations by other threads on the fdp so we have to be 2359 * careful. 2360 */ 2361 newfdp = kmalloc(sizeof(struct filedesc), 2362 M_FILEDESC, M_WAITOK | M_ZERO | M_NULLOK); 2363 if (newfdp == NULL) { 2364 *fpp = NULL; 2365 return (-1); 2366 } 2367 again: 2368 spin_lock(&fdp->fd_spin); 2369 if (fdp->fd_lastfile < NDFILE) { 2370 newfdp->fd_files = newfdp->fd_builtin_files; 2371 i = NDFILE; 2372 } else { 2373 /* 2374 * We have to allocate (N^2-1) entries for our in-place 2375 * binary tree. Allow the table to shrink. 2376 */ 2377 i = fdp->fd_nfiles; 2378 ni = (i - 1) / 2; 2379 while (ni > fdp->fd_lastfile && ni > NDFILE) { 2380 i = ni; 2381 ni = (i - 1) / 2; 2382 } 2383 spin_unlock(&fdp->fd_spin); 2384 newfdp->fd_files = kmalloc(i * sizeof(struct fdnode), 2385 M_FILEDESC, M_WAITOK | M_ZERO); 2386 2387 /* 2388 * Check for race, retry 2389 */ 2390 spin_lock(&fdp->fd_spin); 2391 if (i <= fdp->fd_lastfile) { 2392 spin_unlock(&fdp->fd_spin); 2393 kfree(newfdp->fd_files, M_FILEDESC); 2394 goto again; 2395 } 2396 } 2397 2398 /* 2399 * Dup the remaining fields. vref() and cache_hold() can be 2400 * safely called while holding the read spinlock on fdp. 2401 * 2402 * The read spinlock on fdp is still being held. 2403 * 2404 * NOTE: vref and cache_hold calls for the case where the vnode 2405 * or cache entry already has at least one ref may be called 2406 * while holding spin locks. 2407 */ 2408 if ((newfdp->fd_cdir = fdp->fd_cdir) != NULL) { 2409 vref(newfdp->fd_cdir); 2410 cache_copy(&fdp->fd_ncdir, &newfdp->fd_ncdir); 2411 } 2412 /* 2413 * We must check for fd_rdir here, at least for now because 2414 * the init process is created before we have access to the 2415 * rootvode to take a reference to it. 2416 */ 2417 if ((newfdp->fd_rdir = fdp->fd_rdir) != NULL) { 2418 vref(newfdp->fd_rdir); 2419 cache_copy(&fdp->fd_nrdir, &newfdp->fd_nrdir); 2420 } 2421 if ((newfdp->fd_jdir = fdp->fd_jdir) != NULL) { 2422 vref(newfdp->fd_jdir); 2423 cache_copy(&fdp->fd_njdir, &newfdp->fd_njdir); 2424 } 2425 newfdp->fd_refcnt = 1; 2426 newfdp->fd_nfiles = i; 2427 newfdp->fd_lastfile = fdp->fd_lastfile; 2428 newfdp->fd_freefile = fdp->fd_freefile; 2429 newfdp->fd_cmask = fdp->fd_cmask; 2430 spin_init(&newfdp->fd_spin, "fdcopy"); 2431 2432 /* 2433 * Copy the descriptor table through (i). This also copies the 2434 * allocation state. Then go through and ref the file pointers 2435 * and clean up any KQ descriptors. 2436 * 2437 * kq descriptors cannot be copied. Since we haven't ref'd the 2438 * copied files yet we can ignore the return value from funsetfd(). 2439 * 2440 * The read spinlock on fdp is still being held. 2441 * 2442 * Be sure to clean out fdnode->tdcache, otherwise bad things will 2443 * happen. 2444 */ 2445 bcopy(fdp->fd_files, newfdp->fd_files, i * sizeof(struct fdnode)); 2446 for (i = 0 ; i < newfdp->fd_nfiles; ++i) { 2447 fdnode = &newfdp->fd_files[i]; 2448 if (fdnode->reserved) { 2449 fdreserve_locked(newfdp, i, -1); 2450 fdnode->reserved = 0; 2451 fdfixup_locked(newfdp, i); 2452 } else if (fdnode->fp) { 2453 bzero(&fdnode->tdcache, sizeof(fdnode->tdcache)); 2454 if (fdnode->fp->f_type == DTYPE_KQUEUE) { 2455 (void)funsetfd_locked(newfdp, i); 2456 } else { 2457 fhold(fdnode->fp); 2458 } 2459 } 2460 } 2461 spin_unlock(&fdp->fd_spin); 2462 *fpp = newfdp; 2463 return (0); 2464 } 2465 2466 /* 2467 * Release a filedesc structure. 2468 * 2469 * NOT MPSAFE (MPSAFE for refs > 1, but the final cleanup code is not MPSAFE) 2470 */ 2471 void 2472 fdfree(struct proc *p, struct filedesc *repl) 2473 { 2474 struct filedesc *fdp; 2475 struct fdnode *fdnode; 2476 int i; 2477 struct filedesc_to_leader *fdtol; 2478 struct file *fp; 2479 struct vnode *vp; 2480 struct flock lf; 2481 2482 /* 2483 * Before destroying or replacing p->p_fd we must be sure to 2484 * clean out the cache of the last thread, which should be 2485 * curthread. 2486 */ 2487 fexitcache(curthread); 2488 2489 /* 2490 * Certain daemons might not have file descriptors. 2491 */ 2492 fdp = p->p_fd; 2493 if (fdp == NULL) { 2494 p->p_fd = repl; 2495 return; 2496 } 2497 2498 /* 2499 * Severe messing around to follow. 2500 */ 2501 spin_lock(&fdp->fd_spin); 2502 2503 /* Check for special need to clear POSIX style locks */ 2504 fdtol = p->p_fdtol; 2505 if (fdtol != NULL) { 2506 KASSERT(fdtol->fdl_refcount > 0, 2507 ("filedesc_to_refcount botch: fdl_refcount=%d", 2508 fdtol->fdl_refcount)); 2509 if (fdtol->fdl_refcount == 1 && p->p_leader->p_advlock_flag) { 2510 for (i = 0; i <= fdp->fd_lastfile; ++i) { 2511 fdnode = &fdp->fd_files[i]; 2512 if (fdnode->fp == NULL || 2513 fdnode->fp->f_type != DTYPE_VNODE) { 2514 continue; 2515 } 2516 fp = fdnode->fp; 2517 fhold(fp); 2518 spin_unlock(&fdp->fd_spin); 2519 2520 lf.l_whence = SEEK_SET; 2521 lf.l_start = 0; 2522 lf.l_len = 0; 2523 lf.l_type = F_UNLCK; 2524 vp = (struct vnode *)fp->f_data; 2525 VOP_ADVLOCK(vp, (caddr_t)p->p_leader, 2526 F_UNLCK, &lf, F_POSIX); 2527 fdrop(fp); 2528 spin_lock(&fdp->fd_spin); 2529 } 2530 } 2531 retry: 2532 if (fdtol->fdl_refcount == 1) { 2533 if (fdp->fd_holdleaderscount > 0 && 2534 p->p_leader->p_advlock_flag) { 2535 /* 2536 * close() or do_dup() has cleared a reference 2537 * in a shared file descriptor table. 2538 */ 2539 fdp->fd_holdleaderswakeup = 1; 2540 ssleep(&fdp->fd_holdleaderscount, 2541 &fdp->fd_spin, 0, "fdlhold", 0); 2542 goto retry; 2543 } 2544 if (fdtol->fdl_holdcount > 0) { 2545 /* 2546 * Ensure that fdtol->fdl_leader 2547 * remains valid in closef(). 2548 */ 2549 fdtol->fdl_wakeup = 1; 2550 ssleep(fdtol, &fdp->fd_spin, 0, "fdlhold", 0); 2551 goto retry; 2552 } 2553 } 2554 fdtol->fdl_refcount--; 2555 if (fdtol->fdl_refcount == 0 && 2556 fdtol->fdl_holdcount == 0) { 2557 fdtol->fdl_next->fdl_prev = fdtol->fdl_prev; 2558 fdtol->fdl_prev->fdl_next = fdtol->fdl_next; 2559 } else { 2560 fdtol = NULL; 2561 } 2562 p->p_fdtol = NULL; 2563 if (fdtol != NULL) { 2564 spin_unlock(&fdp->fd_spin); 2565 kfree(fdtol, M_FILEDESC_TO_LEADER); 2566 spin_lock(&fdp->fd_spin); 2567 } 2568 } 2569 if (--fdp->fd_refcnt > 0) { 2570 spin_unlock(&fdp->fd_spin); 2571 spin_lock(&p->p_spin); 2572 p->p_fd = repl; 2573 spin_unlock(&p->p_spin); 2574 return; 2575 } 2576 2577 /* 2578 * Even though we are the last reference to the structure allproc 2579 * scans may still reference the structure. Maintain proper 2580 * locks until we can replace p->p_fd. 2581 * 2582 * Also note that kqueue's closef still needs to reference the 2583 * fdp via p->p_fd, so we have to close the descriptors before 2584 * we replace p->p_fd. 2585 */ 2586 for (i = 0; i <= fdp->fd_lastfile; ++i) { 2587 if (fdp->fd_files[i].fp) { 2588 fp = funsetfd_locked(fdp, i); 2589 if (fp) { 2590 spin_unlock(&fdp->fd_spin); 2591 if (SLIST_FIRST(&fp->f_klist)) 2592 knote_fdclose(fp, fdp, i); 2593 closef(fp, p); 2594 spin_lock(&fdp->fd_spin); 2595 } 2596 } 2597 } 2598 spin_unlock(&fdp->fd_spin); 2599 2600 /* 2601 * Interlock against an allproc scan operations (typically frevoke). 2602 */ 2603 spin_lock(&p->p_spin); 2604 p->p_fd = repl; 2605 spin_unlock(&p->p_spin); 2606 2607 /* 2608 * Wait for any softrefs to go away. This race rarely occurs so 2609 * we can use a non-critical-path style poll/sleep loop. The 2610 * race only occurs against allproc scans. 2611 * 2612 * No new softrefs can occur with the fdp disconnected from the 2613 * process. 2614 */ 2615 if (fdp->fd_softrefs) { 2616 kprintf("pid %d: Warning, fdp race avoided\n", p->p_pid); 2617 while (fdp->fd_softrefs) 2618 tsleep(&fdp->fd_softrefs, 0, "fdsoft", 1); 2619 } 2620 2621 if (fdp->fd_files != fdp->fd_builtin_files) 2622 kfree(fdp->fd_files, M_FILEDESC); 2623 if (fdp->fd_cdir) { 2624 cache_drop(&fdp->fd_ncdir); 2625 vrele(fdp->fd_cdir); 2626 } 2627 if (fdp->fd_rdir) { 2628 cache_drop(&fdp->fd_nrdir); 2629 vrele(fdp->fd_rdir); 2630 } 2631 if (fdp->fd_jdir) { 2632 cache_drop(&fdp->fd_njdir); 2633 vrele(fdp->fd_jdir); 2634 } 2635 kfree(fdp, M_FILEDESC); 2636 } 2637 2638 /* 2639 * Retrieve and reference the file pointer associated with a descriptor. 2640 * 2641 * td must be the current thread. 2642 */ 2643 struct file * 2644 holdfp(thread_t td, int fd, int flag) 2645 { 2646 struct file *fp; 2647 2648 fp = _holdfp_cache(td, fd); 2649 if (fp) { 2650 if ((fp->f_flag & flag) == 0 && flag != -1) { 2651 fdrop(fp); 2652 fp = NULL; 2653 } 2654 } 2655 return fp; 2656 } 2657 2658 /* 2659 * holdsock() - load the struct file pointer associated 2660 * with a socket into *fpp. If an error occurs, non-zero 2661 * will be returned and *fpp will be set to NULL. 2662 * 2663 * td must be the current thread. 2664 */ 2665 int 2666 holdsock(thread_t td, int fd, struct file **fpp) 2667 { 2668 struct file *fp; 2669 int error; 2670 2671 /* 2672 * Lockless shortcut 2673 */ 2674 fp = _holdfp_cache(td, fd); 2675 if (fp) { 2676 if (fp->f_type != DTYPE_SOCKET) { 2677 fdrop(fp); 2678 fp = NULL; 2679 error = ENOTSOCK; 2680 } else { 2681 error = 0; 2682 } 2683 } else { 2684 error = EBADF; 2685 } 2686 *fpp = fp; 2687 2688 return (error); 2689 } 2690 2691 /* 2692 * Convert a user file descriptor to a held file pointer. 2693 * 2694 * td must be the current thread. 2695 */ 2696 int 2697 holdvnode(thread_t td, int fd, struct file **fpp) 2698 { 2699 struct file *fp; 2700 int error; 2701 2702 fp = _holdfp_cache(td, fd); 2703 if (fp) { 2704 if (fp->f_type != DTYPE_VNODE && fp->f_type != DTYPE_FIFO) { 2705 fdrop(fp); 2706 fp = NULL; 2707 error = EINVAL; 2708 } else { 2709 error = 0; 2710 } 2711 } else { 2712 error = EBADF; 2713 } 2714 *fpp = fp; 2715 2716 return (error); 2717 } 2718 2719 /* 2720 * For setugid programs, we don't want to people to use that setugidness 2721 * to generate error messages which write to a file which otherwise would 2722 * otherwise be off-limits to the process. 2723 * 2724 * This is a gross hack to plug the hole. A better solution would involve 2725 * a special vop or other form of generalized access control mechanism. We 2726 * go ahead and just reject all procfs file systems accesses as dangerous. 2727 * 2728 * Since setugidsafety calls this only for fd 0, 1 and 2, this check is 2729 * sufficient. We also don't for check setugidness since we know we are. 2730 */ 2731 static int 2732 is_unsafe(struct file *fp) 2733 { 2734 if (fp->f_type == DTYPE_VNODE && 2735 ((struct vnode *)(fp->f_data))->v_tag == VT_PROCFS) 2736 return (1); 2737 return (0); 2738 } 2739 2740 /* 2741 * Make this setguid thing safe, if at all possible. 2742 * 2743 * NOT MPSAFE - scans fdp without spinlocks, calls knote_fdclose() 2744 */ 2745 void 2746 setugidsafety(struct proc *p) 2747 { 2748 struct filedesc *fdp = p->p_fd; 2749 int i; 2750 2751 /* Certain daemons might not have file descriptors. */ 2752 if (fdp == NULL) 2753 return; 2754 2755 /* 2756 * note: fdp->fd_files may be reallocated out from under us while 2757 * we are blocked in a close. Be careful! 2758 */ 2759 for (i = 0; i <= fdp->fd_lastfile; i++) { 2760 if (i > 2) 2761 break; 2762 if (fdp->fd_files[i].fp && is_unsafe(fdp->fd_files[i].fp)) { 2763 struct file *fp; 2764 2765 /* 2766 * NULL-out descriptor prior to close to avoid 2767 * a race while close blocks. 2768 */ 2769 if ((fp = funsetfd_locked(fdp, i)) != NULL) { 2770 knote_fdclose(fp, fdp, i); 2771 closef(fp, p); 2772 } 2773 } 2774 } 2775 } 2776 2777 /* 2778 * Close all CLOEXEC files on exec. 2779 * 2780 * Only a single thread remains for the current process. 2781 * 2782 * NOT MPSAFE - scans fdp without spinlocks, calls knote_fdclose() 2783 */ 2784 void 2785 fdcloseexec(struct proc *p) 2786 { 2787 struct filedesc *fdp = p->p_fd; 2788 int i; 2789 2790 /* Certain daemons might not have file descriptors. */ 2791 if (fdp == NULL) 2792 return; 2793 2794 /* 2795 * We cannot cache fd_files since operations may block and rip 2796 * them out from under us. 2797 */ 2798 for (i = 0; i <= fdp->fd_lastfile; i++) { 2799 if (fdp->fd_files[i].fp != NULL && 2800 (fdp->fd_files[i].fileflags & UF_EXCLOSE)) { 2801 struct file *fp; 2802 2803 /* 2804 * NULL-out descriptor prior to close to avoid 2805 * a race while close blocks. 2806 * 2807 * (funsetfd*() also clears the fd cache) 2808 */ 2809 if ((fp = funsetfd_locked(fdp, i)) != NULL) { 2810 knote_fdclose(fp, fdp, i); 2811 closef(fp, p); 2812 } 2813 } 2814 } 2815 } 2816 2817 /* 2818 * It is unsafe for set[ug]id processes to be started with file 2819 * descriptors 0..2 closed, as these descriptors are given implicit 2820 * significance in the Standard C library. fdcheckstd() will create a 2821 * descriptor referencing /dev/null for each of stdin, stdout, and 2822 * stderr that is not already open. 2823 * 2824 * NOT MPSAFE - calls falloc, vn_open, etc 2825 */ 2826 int 2827 fdcheckstd(struct lwp *lp) 2828 { 2829 struct nlookupdata nd; 2830 struct filedesc *fdp; 2831 struct file *fp; 2832 int retval; 2833 int i, error, flags, devnull; 2834 2835 fdp = lp->lwp_proc->p_fd; 2836 if (fdp == NULL) 2837 return (0); 2838 devnull = -1; 2839 error = 0; 2840 for (i = 0; i < 3; i++) { 2841 if (fdp->fd_files[i].fp != NULL) 2842 continue; 2843 if (devnull < 0) { 2844 if ((error = falloc(lp, &fp, &devnull)) != 0) 2845 break; 2846 2847 error = nlookup_init(&nd, "/dev/null", UIO_SYSSPACE, 2848 NLC_FOLLOW|NLC_LOCKVP); 2849 flags = FREAD | FWRITE; 2850 if (error == 0) 2851 error = vn_open(&nd, fp, flags, 0); 2852 if (error == 0) 2853 fsetfd(fdp, fp, devnull); 2854 else 2855 fsetfd(fdp, NULL, devnull); 2856 fdrop(fp); 2857 nlookup_done(&nd); 2858 if (error) 2859 break; 2860 KKASSERT(i == devnull); 2861 } else { 2862 error = kern_dup(DUP_FIXED, devnull, i, &retval); 2863 if (error != 0) 2864 break; 2865 } 2866 } 2867 return (error); 2868 } 2869 2870 /* 2871 * Internal form of close. 2872 * Decrement reference count on file structure. 2873 * Note: td and/or p may be NULL when closing a file 2874 * that was being passed in a message. 2875 * 2876 * MPALMOSTSAFE - acquires mplock for VOP operations 2877 */ 2878 int 2879 closef(struct file *fp, struct proc *p) 2880 { 2881 struct vnode *vp; 2882 struct flock lf; 2883 struct filedesc_to_leader *fdtol; 2884 2885 if (fp == NULL) 2886 return (0); 2887 2888 /* 2889 * POSIX record locking dictates that any close releases ALL 2890 * locks owned by this process. This is handled by setting 2891 * a flag in the unlock to free ONLY locks obeying POSIX 2892 * semantics, and not to free BSD-style file locks. 2893 * If the descriptor was in a message, POSIX-style locks 2894 * aren't passed with the descriptor. 2895 */ 2896 if (p != NULL && fp->f_type == DTYPE_VNODE && 2897 (((struct vnode *)fp->f_data)->v_flag & VMAYHAVELOCKS) 2898 ) { 2899 if (p->p_leader->p_advlock_flag) { 2900 lf.l_whence = SEEK_SET; 2901 lf.l_start = 0; 2902 lf.l_len = 0; 2903 lf.l_type = F_UNLCK; 2904 vp = (struct vnode *)fp->f_data; 2905 VOP_ADVLOCK(vp, (caddr_t)p->p_leader, F_UNLCK, 2906 &lf, F_POSIX); 2907 } 2908 fdtol = p->p_fdtol; 2909 if (fdtol != NULL) { 2910 lwkt_gettoken(&p->p_token); 2911 2912 /* 2913 * Handle special case where file descriptor table 2914 * is shared between multiple process leaders. 2915 */ 2916 for (fdtol = fdtol->fdl_next; 2917 fdtol != p->p_fdtol; 2918 fdtol = fdtol->fdl_next) { 2919 if (fdtol->fdl_leader->p_advlock_flag == 0) 2920 continue; 2921 fdtol->fdl_holdcount++; 2922 lf.l_whence = SEEK_SET; 2923 lf.l_start = 0; 2924 lf.l_len = 0; 2925 lf.l_type = F_UNLCK; 2926 vp = (struct vnode *)fp->f_data; 2927 VOP_ADVLOCK(vp, (caddr_t)fdtol->fdl_leader, 2928 F_UNLCK, &lf, F_POSIX); 2929 fdtol->fdl_holdcount--; 2930 if (fdtol->fdl_holdcount == 0 && 2931 fdtol->fdl_wakeup != 0) { 2932 fdtol->fdl_wakeup = 0; 2933 wakeup(fdtol); 2934 } 2935 } 2936 lwkt_reltoken(&p->p_token); 2937 } 2938 } 2939 return (fdrop(fp)); 2940 } 2941 2942 /* 2943 * fhold() can only be called if f_count is already at least 1 (i.e. the 2944 * caller of fhold() already has a reference to the file pointer in some 2945 * manner or other). 2946 * 2947 * Atomic ops are used for incrementing and decrementing f_count before 2948 * the 1->0 transition. f_count 1->0 transition is special, see the 2949 * comment in fdrop(). 2950 */ 2951 void 2952 fhold(struct file *fp) 2953 { 2954 /* 0->1 transition will never work */ 2955 KASSERT(fp->f_count > 0, ("fhold: invalid f_count %d", fp->f_count)); 2956 atomic_add_int(&fp->f_count, 1); 2957 } 2958 2959 /* 2960 * fdrop() - drop a reference to a descriptor 2961 */ 2962 int 2963 fdrop(struct file *fp) 2964 { 2965 struct flock lf; 2966 struct vnode *vp; 2967 int error, do_free = 0; 2968 2969 /* 2970 * NOTE: 2971 * Simple atomic_fetchadd_int(f_count, -1) here will cause use- 2972 * after-free or double free (due to f_count 0->1 transition), if 2973 * fhold() is called on the fps found through filehead iteration. 2974 */ 2975 for (;;) { 2976 int count = fp->f_count; 2977 2978 cpu_ccfence(); 2979 KASSERT(count > 0, ("fdrop: invalid f_count %d", count)); 2980 if (count == 1) { 2981 struct filelist_head *head = fp2filelist(fp); 2982 2983 /* 2984 * About to drop the last reference, hold the 2985 * filehead spin lock and drop it, so that no 2986 * one could see this fp through filehead anymore, 2987 * let alone fhold() this fp. 2988 */ 2989 spin_lock(&head->spin); 2990 if (atomic_cmpset_int(&fp->f_count, count, 0)) { 2991 LIST_REMOVE(fp, f_list); 2992 spin_unlock(&head->spin); 2993 atomic_subtract_int(&nfiles, 1); 2994 do_free = 1; /* free this fp */ 2995 break; 2996 } 2997 spin_unlock(&head->spin); 2998 /* retry */ 2999 } else if (atomic_cmpset_int(&fp->f_count, count, count - 1)) { 3000 break; 3001 } 3002 /* retry */ 3003 } 3004 if (!do_free) 3005 return (0); 3006 3007 KKASSERT(SLIST_FIRST(&fp->f_klist) == NULL); 3008 3009 /* 3010 * The last reference has gone away, we own the fp structure free 3011 * and clear. 3012 */ 3013 if (fp->f_count < 0) 3014 panic("fdrop: count < 0"); 3015 if ((fp->f_flag & FHASLOCK) && fp->f_type == DTYPE_VNODE && 3016 (((struct vnode *)fp->f_data)->v_flag & VMAYHAVELOCKS) 3017 ) { 3018 lf.l_whence = SEEK_SET; 3019 lf.l_start = 0; 3020 lf.l_len = 0; 3021 lf.l_type = F_UNLCK; 3022 vp = (struct vnode *)fp->f_data; 3023 VOP_ADVLOCK(vp, (caddr_t)fp, F_UNLCK, &lf, 0); 3024 } 3025 if (fp->f_ops != &badfileops) 3026 error = fo_close(fp); 3027 else 3028 error = 0; 3029 ffree(fp); 3030 return (error); 3031 } 3032 3033 /* 3034 * Apply an advisory lock on a file descriptor. 3035 * 3036 * Just attempt to get a record lock of the requested type on 3037 * the entire file (l_whence = SEEK_SET, l_start = 0, l_len = 0). 3038 * 3039 * MPALMOSTSAFE 3040 */ 3041 int 3042 sys_flock(struct flock_args *uap) 3043 { 3044 thread_t td = curthread; 3045 struct file *fp; 3046 struct vnode *vp; 3047 struct flock lf; 3048 int error; 3049 3050 if ((fp = holdfp(td, uap->fd, -1)) == NULL) 3051 return (EBADF); 3052 if (fp->f_type != DTYPE_VNODE) { 3053 error = EOPNOTSUPP; 3054 goto done; 3055 } 3056 vp = (struct vnode *)fp->f_data; 3057 lf.l_whence = SEEK_SET; 3058 lf.l_start = 0; 3059 lf.l_len = 0; 3060 if (uap->how & LOCK_UN) { 3061 lf.l_type = F_UNLCK; 3062 atomic_clear_int(&fp->f_flag, FHASLOCK); /* race ok */ 3063 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_UNLCK, &lf, 0); 3064 goto done; 3065 } 3066 if (uap->how & LOCK_EX) 3067 lf.l_type = F_WRLCK; 3068 else if (uap->how & LOCK_SH) 3069 lf.l_type = F_RDLCK; 3070 else { 3071 error = EBADF; 3072 goto done; 3073 } 3074 if (uap->how & LOCK_NB) 3075 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, 0); 3076 else 3077 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, F_WAIT); 3078 atomic_set_int(&fp->f_flag, FHASLOCK); /* race ok */ 3079 done: 3080 fdrop(fp); 3081 return (error); 3082 } 3083 3084 /* 3085 * File Descriptor pseudo-device driver (/dev/fd/). 3086 * 3087 * Opening minor device N dup()s the file (if any) connected to file 3088 * descriptor N belonging to the calling process. Note that this driver 3089 * consists of only the ``open()'' routine, because all subsequent 3090 * references to this file will be direct to the other driver. 3091 */ 3092 static int 3093 fdopen(struct dev_open_args *ap) 3094 { 3095 thread_t td = curthread; 3096 3097 KKASSERT(td->td_lwp != NULL); 3098 3099 /* 3100 * XXX Kludge: set curlwp->lwp_dupfd to contain the value of the 3101 * the file descriptor being sought for duplication. The error 3102 * return ensures that the vnode for this device will be released 3103 * by vn_open. Open will detect this special error and take the 3104 * actions in dupfdopen below. Other callers of vn_open or VOP_OPEN 3105 * will simply report the error. 3106 */ 3107 td->td_lwp->lwp_dupfd = minor(ap->a_head.a_dev); 3108 return (ENODEV); 3109 } 3110 3111 /* 3112 * The caller has reserved the file descriptor dfd for us. On success we 3113 * must fsetfd() it. On failure the caller will clean it up. 3114 */ 3115 int 3116 dupfdopen(thread_t td, int dfd, int sfd, int mode, int error) 3117 { 3118 struct filedesc *fdp; 3119 struct file *wfp; 3120 struct file *xfp; 3121 int werror; 3122 3123 if ((wfp = holdfp(td, sfd, -1)) == NULL) 3124 return (EBADF); 3125 3126 /* 3127 * Close a revoke/dup race. Duping a descriptor marked as revoked 3128 * will dup a dummy descriptor instead of the real one. 3129 */ 3130 if (wfp->f_flag & FREVOKED) { 3131 kprintf("Warning: attempt to dup() a revoked descriptor\n"); 3132 fdrop(wfp); 3133 wfp = NULL; 3134 werror = falloc(NULL, &wfp, NULL); 3135 if (werror) 3136 return (werror); 3137 } 3138 3139 fdp = td->td_proc->p_fd; 3140 3141 /* 3142 * There are two cases of interest here. 3143 * 3144 * For ENODEV simply dup sfd to file descriptor dfd and return. 3145 * 3146 * For ENXIO steal away the file structure from sfd and store it 3147 * dfd. sfd is effectively closed by this operation. 3148 * 3149 * Any other error code is just returned. 3150 */ 3151 switch (error) { 3152 case ENODEV: 3153 /* 3154 * Check that the mode the file is being opened for is a 3155 * subset of the mode of the existing descriptor. 3156 */ 3157 if (((mode & (FREAD|FWRITE)) | wfp->f_flag) != wfp->f_flag) { 3158 error = EACCES; 3159 break; 3160 } 3161 spin_lock(&fdp->fd_spin); 3162 fdp->fd_files[dfd].fileflags = fdp->fd_files[sfd].fileflags; 3163 fsetfd_locked(fdp, wfp, dfd); 3164 spin_unlock(&fdp->fd_spin); 3165 error = 0; 3166 break; 3167 case ENXIO: 3168 /* 3169 * Steal away the file pointer from dfd, and stuff it into indx. 3170 */ 3171 spin_lock(&fdp->fd_spin); 3172 fdp->fd_files[dfd].fileflags = fdp->fd_files[sfd].fileflags; 3173 fsetfd(fdp, wfp, dfd); 3174 if ((xfp = funsetfd_locked(fdp, sfd)) != NULL) { 3175 spin_unlock(&fdp->fd_spin); 3176 fdrop(xfp); 3177 } else { 3178 spin_unlock(&fdp->fd_spin); 3179 } 3180 error = 0; 3181 break; 3182 default: 3183 break; 3184 } 3185 fdrop(wfp); 3186 return (error); 3187 } 3188 3189 /* 3190 * NOT MPSAFE - I think these refer to a common file descriptor table 3191 * and we need to spinlock that to link fdtol in. 3192 */ 3193 struct filedesc_to_leader * 3194 filedesc_to_leader_alloc(struct filedesc_to_leader *old, 3195 struct proc *leader) 3196 { 3197 struct filedesc_to_leader *fdtol; 3198 3199 fdtol = kmalloc(sizeof(struct filedesc_to_leader), 3200 M_FILEDESC_TO_LEADER, M_WAITOK | M_ZERO); 3201 fdtol->fdl_refcount = 1; 3202 fdtol->fdl_holdcount = 0; 3203 fdtol->fdl_wakeup = 0; 3204 fdtol->fdl_leader = leader; 3205 if (old != NULL) { 3206 fdtol->fdl_next = old->fdl_next; 3207 fdtol->fdl_prev = old; 3208 old->fdl_next = fdtol; 3209 fdtol->fdl_next->fdl_prev = fdtol; 3210 } else { 3211 fdtol->fdl_next = fdtol; 3212 fdtol->fdl_prev = fdtol; 3213 } 3214 return fdtol; 3215 } 3216 3217 /* 3218 * Scan all file pointers in the system. The callback is made with 3219 * the master list spinlock held exclusively. 3220 */ 3221 void 3222 allfiles_scan_exclusive(int (*callback)(struct file *, void *), void *data) 3223 { 3224 int i; 3225 3226 for (i = 0; i < NFILELIST_HEADS; ++i) { 3227 struct filelist_head *head = &filelist_heads[i]; 3228 struct file *fp; 3229 3230 spin_lock(&head->spin); 3231 LIST_FOREACH(fp, &head->list, f_list) { 3232 int res; 3233 3234 res = callback(fp, data); 3235 if (res < 0) 3236 break; 3237 } 3238 spin_unlock(&head->spin); 3239 } 3240 } 3241 3242 /* 3243 * Get file structures. 3244 * 3245 * NOT MPSAFE - process list scan, SYSCTL_OUT (probably not mpsafe) 3246 */ 3247 3248 struct sysctl_kern_file_info { 3249 int count; 3250 int error; 3251 struct sysctl_req *req; 3252 }; 3253 3254 static int sysctl_kern_file_callback(struct proc *p, void *data); 3255 3256 static int 3257 sysctl_kern_file(SYSCTL_HANDLER_ARGS) 3258 { 3259 struct sysctl_kern_file_info info; 3260 3261 /* 3262 * Note: because the number of file descriptors is calculated 3263 * in different ways for sizing vs returning the data, 3264 * there is information leakage from the first loop. However, 3265 * it is of a similar order of magnitude to the leakage from 3266 * global system statistics such as kern.openfiles. 3267 * 3268 * When just doing a count, note that we cannot just count 3269 * the elements and add f_count via the filehead list because 3270 * threaded processes share their descriptor table and f_count might 3271 * still be '1' in that case. 3272 * 3273 * Since the SYSCTL op can block, we must hold the process to 3274 * prevent it being ripped out from under us either in the 3275 * file descriptor loop or in the greater LIST_FOREACH. The 3276 * process may be in varying states of disrepair. If the process 3277 * is in SZOMB we may have caught it just as it is being removed 3278 * from the allproc list, we must skip it in that case to maintain 3279 * an unbroken chain through the allproc list. 3280 */ 3281 info.count = 0; 3282 info.error = 0; 3283 info.req = req; 3284 allproc_scan(sysctl_kern_file_callback, &info, 0); 3285 3286 /* 3287 * When just calculating the size, overestimate a bit to try to 3288 * prevent system activity from causing the buffer-fill call 3289 * to fail later on. 3290 */ 3291 if (req->oldptr == NULL) { 3292 info.count = (info.count + 16) + (info.count / 10); 3293 info.error = SYSCTL_OUT(req, NULL, 3294 info.count * sizeof(struct kinfo_file)); 3295 } 3296 return (info.error); 3297 } 3298 3299 static int 3300 sysctl_kern_file_callback(struct proc *p, void *data) 3301 { 3302 struct sysctl_kern_file_info *info = data; 3303 struct kinfo_file kf; 3304 struct filedesc *fdp; 3305 struct file *fp; 3306 uid_t uid; 3307 int n; 3308 3309 if (p->p_stat == SIDL || p->p_stat == SZOMB) 3310 return(0); 3311 if (!(PRISON_CHECK(info->req->td->td_ucred, p->p_ucred) != 0)) 3312 return(0); 3313 3314 /* 3315 * Softref the fdp to prevent it from being destroyed 3316 */ 3317 spin_lock(&p->p_spin); 3318 if ((fdp = p->p_fd) == NULL) { 3319 spin_unlock(&p->p_spin); 3320 return(0); 3321 } 3322 atomic_add_int(&fdp->fd_softrefs, 1); 3323 spin_unlock(&p->p_spin); 3324 3325 /* 3326 * The fdp's own spinlock prevents the contents from being 3327 * modified. 3328 */ 3329 spin_lock_shared(&fdp->fd_spin); 3330 for (n = 0; n < fdp->fd_nfiles; ++n) { 3331 if ((fp = fdp->fd_files[n].fp) == NULL) 3332 continue; 3333 if (info->req->oldptr == NULL) { 3334 ++info->count; 3335 } else { 3336 uid = p->p_ucred ? p->p_ucred->cr_uid : -1; 3337 kcore_make_file(&kf, fp, p->p_pid, uid, n); 3338 spin_unlock_shared(&fdp->fd_spin); 3339 info->error = SYSCTL_OUT(info->req, &kf, sizeof(kf)); 3340 spin_lock_shared(&fdp->fd_spin); 3341 if (info->error) 3342 break; 3343 } 3344 } 3345 spin_unlock_shared(&fdp->fd_spin); 3346 atomic_subtract_int(&fdp->fd_softrefs, 1); 3347 if (info->error) 3348 return(-1); 3349 return(0); 3350 } 3351 3352 SYSCTL_PROC(_kern, KERN_FILE, file, CTLTYPE_OPAQUE|CTLFLAG_RD, 3353 0, 0, sysctl_kern_file, "S,file", "Entire file table"); 3354 3355 SYSCTL_INT(_kern, OID_AUTO, minfilesperproc, CTLFLAG_RW, 3356 &minfilesperproc, 0, "Minimum files allowed open per process"); 3357 SYSCTL_INT(_kern, KERN_MAXFILESPERPROC, maxfilesperproc, CTLFLAG_RW, 3358 &maxfilesperproc, 0, "Maximum files allowed open per process"); 3359 SYSCTL_INT(_kern, OID_AUTO, maxfilesperuser, CTLFLAG_RW, 3360 &maxfilesperuser, 0, "Maximum files allowed open per user"); 3361 3362 SYSCTL_INT(_kern, KERN_MAXFILES, maxfiles, CTLFLAG_RW, 3363 &maxfiles, 0, "Maximum number of files"); 3364 3365 SYSCTL_INT(_kern, OID_AUTO, maxfilesrootres, CTLFLAG_RW, 3366 &maxfilesrootres, 0, "Descriptors reserved for root use"); 3367 3368 SYSCTL_INT(_kern, OID_AUTO, openfiles, CTLFLAG_RD, 3369 &nfiles, 0, "System-wide number of open files"); 3370 3371 static void 3372 fildesc_drvinit(void *unused) 3373 { 3374 int fd; 3375 3376 for (fd = 0; fd < NUMFDESC; fd++) { 3377 make_dev(&fildesc_ops, fd, 3378 UID_BIN, GID_BIN, 0666, "fd/%d", fd); 3379 } 3380 3381 make_dev(&fildesc_ops, 0, UID_ROOT, GID_WHEEL, 0666, "stdin"); 3382 make_dev(&fildesc_ops, 1, UID_ROOT, GID_WHEEL, 0666, "stdout"); 3383 make_dev(&fildesc_ops, 2, UID_ROOT, GID_WHEEL, 0666, "stderr"); 3384 } 3385 3386 struct fileops badfileops = { 3387 .fo_read = badfo_readwrite, 3388 .fo_write = badfo_readwrite, 3389 .fo_ioctl = badfo_ioctl, 3390 .fo_kqfilter = badfo_kqfilter, 3391 .fo_stat = badfo_stat, 3392 .fo_close = badfo_close, 3393 .fo_shutdown = badfo_shutdown 3394 }; 3395 3396 int 3397 badfo_readwrite( 3398 struct file *fp, 3399 struct uio *uio, 3400 struct ucred *cred, 3401 int flags 3402 ) { 3403 return (EBADF); 3404 } 3405 3406 int 3407 badfo_ioctl(struct file *fp, u_long com, caddr_t data, 3408 struct ucred *cred, struct sysmsg *msgv) 3409 { 3410 return (EBADF); 3411 } 3412 3413 /* 3414 * Must return an error to prevent registration, typically 3415 * due to a revoked descriptor (file_filtops assigned). 3416 */ 3417 int 3418 badfo_kqfilter(struct file *fp, struct knote *kn) 3419 { 3420 return (EOPNOTSUPP); 3421 } 3422 3423 int 3424 badfo_stat(struct file *fp, struct stat *sb, struct ucred *cred) 3425 { 3426 return (EBADF); 3427 } 3428 3429 int 3430 badfo_close(struct file *fp) 3431 { 3432 return (EBADF); 3433 } 3434 3435 int 3436 badfo_shutdown(struct file *fp, int how) 3437 { 3438 return (EBADF); 3439 } 3440 3441 int 3442 nofo_shutdown(struct file *fp, int how) 3443 { 3444 return (EOPNOTSUPP); 3445 } 3446 3447 SYSINIT(fildescdev, SI_SUB_DRIVERS, SI_ORDER_MIDDLE + CDEV_MAJOR, 3448 fildesc_drvinit,NULL); 3449 3450 static void 3451 filelist_heads_init(void *arg __unused) 3452 { 3453 int i; 3454 3455 for (i = 0; i < NFILELIST_HEADS; ++i) { 3456 struct filelist_head *head = &filelist_heads[i]; 3457 3458 spin_init(&head->spin, "filehead_spin"); 3459 LIST_INIT(&head->list); 3460 } 3461 } 3462 3463 SYSINIT(filelistheads, SI_BOOT1_LOCK, SI_ORDER_ANY, 3464 filelist_heads_init, NULL); 3465 3466 static void 3467 file_objcache_init(void *dummy __unused) 3468 { 3469 file_objcache = objcache_create("file", maxfiles, maxfiles / 8, 3470 NULL, NULL, NULL, /* TODO: ctor/dtor */ 3471 objcache_malloc_alloc, objcache_malloc_free, &file_malloc_args); 3472 } 3473 SYSINIT(fpobjcache, SI_BOOT2_POST_SMP, SI_ORDER_ANY, file_objcache_init, NULL); 3474