kern_descrip.c revision 1.260
1 /* $NetBSD: kern_descrip.c,v 1.260 2023/09/12 16:17:21 ad Exp $ */ 2 3 /*- 4 * Copyright (c) 2008, 2009 The NetBSD Foundation, Inc. 5 * All rights reserved. 6 * 7 * This code is derived from software contributed to The NetBSD Foundation 8 * by Andrew Doran. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 19 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 20 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 21 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 22 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 23 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 24 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 25 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 26 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 27 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 28 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 29 * POSSIBILITY OF SUCH DAMAGE. 30 */ 31 32 /* 33 * Copyright (c) 1982, 1986, 1989, 1991, 1993 34 * The Regents of the University of California. All rights reserved. 35 * (c) UNIX System Laboratories, Inc. 36 * All or some portions of this file are derived from material licensed 37 * to the University of California by American Telephone and Telegraph 38 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 39 * the permission of UNIX System Laboratories, Inc. 40 * 41 * Redistribution and use in source and binary forms, with or without 42 * modification, are permitted provided that the following conditions 43 * are met: 44 * 1. Redistributions of source code must retain the above copyright 45 * notice, this list of conditions and the following disclaimer. 46 * 2. Redistributions in binary form must reproduce the above copyright 47 * notice, this list of conditions and the following disclaimer in the 48 * documentation and/or other materials provided with the distribution. 49 * 3. Neither the name of the University nor the names of its contributors 50 * may be used to endorse or promote products derived from this software 51 * without specific prior written permission. 52 * 53 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 54 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 55 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 56 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 57 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 58 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 59 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 60 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 61 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 62 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 63 * SUCH DAMAGE. 64 * 65 * @(#)kern_descrip.c 8.8 (Berkeley) 2/14/95 66 */ 67 68 /* 69 * File descriptor management. 70 */ 71 72 #include <sys/cdefs.h> 73 __KERNEL_RCSID(0, "$NetBSD: kern_descrip.c,v 1.260 2023/09/12 16:17:21 ad Exp $"); 74 75 #include <sys/param.h> 76 #include <sys/systm.h> 77 #include <sys/filedesc.h> 78 #include <sys/kernel.h> 79 #include <sys/proc.h> 80 #include <sys/file.h> 81 #include <sys/socket.h> 82 #include <sys/socketvar.h> 83 #include <sys/stat.h> 84 #include <sys/ioctl.h> 85 #include <sys/fcntl.h> 86 #include <sys/pool.h> 87 #include <sys/unistd.h> 88 #include <sys/resourcevar.h> 89 #include <sys/conf.h> 90 #include <sys/event.h> 91 #include <sys/kauth.h> 92 #include <sys/atomic.h> 93 #include <sys/syscallargs.h> 94 #include <sys/cpu.h> 95 #include <sys/kmem.h> 96 #include <sys/vnode.h> 97 #include <sys/sysctl.h> 98 #include <sys/ktrace.h> 99 100 /* 101 * A list (head) of open files, counter, and lock protecting them. 102 */ 103 struct filelist filehead __cacheline_aligned; 104 static u_int nfiles __cacheline_aligned; 105 kmutex_t filelist_lock __cacheline_aligned; 106 107 static pool_cache_t filedesc_cache __read_mostly; 108 static pool_cache_t file_cache __read_mostly; 109 static pool_cache_t fdfile_cache __read_mostly; 110 111 static int file_ctor(void *, void *, int); 112 static void file_dtor(void *, void *); 113 static int fdfile_ctor(void *, void *, int); 114 static void fdfile_dtor(void *, void *); 115 static int filedesc_ctor(void *, void *, int); 116 static void filedesc_dtor(void *, void *); 117 static int filedescopen(dev_t, int, int, lwp_t *); 118 119 static int sysctl_kern_file(SYSCTLFN_PROTO); 120 static int sysctl_kern_file2(SYSCTLFN_PROTO); 121 static void fill_file(struct file *, const struct file *); 122 static void fill_file2(struct kinfo_file *, const file_t *, const fdfile_t *, 123 int, pid_t); 124 125 const struct cdevsw filedesc_cdevsw = { 126 .d_open = filedescopen, 127 .d_close = noclose, 128 .d_read = noread, 129 .d_write = nowrite, 130 .d_ioctl = noioctl, 131 .d_stop = nostop, 132 .d_tty = notty, 133 .d_poll = nopoll, 134 .d_mmap = nommap, 135 .d_kqfilter = nokqfilter, 136 .d_discard = nodiscard, 137 .d_flag = D_OTHER | D_MPSAFE 138 }; 139 140 /* For ease of reading. */ 141 __strong_alias(fd_putvnode,fd_putfile) 142 __strong_alias(fd_putsock,fd_putfile) 143 144 /* 145 * Initialize the descriptor system. 146 */ 147 void 148 fd_sys_init(void) 149 { 150 static struct sysctllog *clog; 151 152 mutex_init(&filelist_lock, MUTEX_DEFAULT, IPL_NONE); 153 154 LIST_INIT(&filehead); 155 156 file_cache = pool_cache_init(sizeof(file_t), coherency_unit, 0, 157 0, "file", NULL, IPL_NONE, file_ctor, file_dtor, NULL); 158 KASSERT(file_cache != NULL); 159 160 fdfile_cache = pool_cache_init(sizeof(fdfile_t), coherency_unit, 0, 161 PR_LARGECACHE, "fdfile", NULL, IPL_NONE, fdfile_ctor, fdfile_dtor, 162 NULL); 163 KASSERT(fdfile_cache != NULL); 164 165 filedesc_cache = pool_cache_init(sizeof(filedesc_t), coherency_unit, 166 0, 0, "filedesc", NULL, IPL_NONE, filedesc_ctor, filedesc_dtor, 167 NULL); 168 KASSERT(filedesc_cache != NULL); 169 170 sysctl_createv(&clog, 0, NULL, NULL, 171 CTLFLAG_PERMANENT, 172 CTLTYPE_STRUCT, "file", 173 SYSCTL_DESCR("System open file table"), 174 sysctl_kern_file, 0, NULL, 0, 175 CTL_KERN, KERN_FILE, CTL_EOL); 176 sysctl_createv(&clog, 0, NULL, NULL, 177 CTLFLAG_PERMANENT, 178 CTLTYPE_STRUCT, "file2", 179 SYSCTL_DESCR("System open file table"), 180 sysctl_kern_file2, 0, NULL, 0, 181 CTL_KERN, KERN_FILE2, CTL_EOL); 182 } 183 184 static bool 185 fd_isused(filedesc_t *fdp, unsigned fd) 186 { 187 u_int off = fd >> NDENTRYSHIFT; 188 189 KASSERT(fd < atomic_load_consume(&fdp->fd_dt)->dt_nfiles); 190 191 return (fdp->fd_lomap[off] & (1U << (fd & NDENTRYMASK))) != 0; 192 } 193 194 /* 195 * Verify that the bitmaps match the descriptor table. 196 */ 197 static inline void 198 fd_checkmaps(filedesc_t *fdp) 199 { 200 #ifdef DEBUG 201 fdtab_t *dt; 202 u_int fd; 203 204 KASSERT(fdp->fd_refcnt <= 1 || mutex_owned(&fdp->fd_lock)); 205 206 dt = fdp->fd_dt; 207 if (fdp->fd_refcnt == -1) { 208 /* 209 * fd_free tears down the table without maintaining its bitmap. 210 */ 211 return; 212 } 213 for (fd = 0; fd < dt->dt_nfiles; fd++) { 214 if (fd < NDFDFILE) { 215 KASSERT(dt->dt_ff[fd] == 216 (fdfile_t *)fdp->fd_dfdfile[fd]); 217 } 218 if (dt->dt_ff[fd] == NULL) { 219 KASSERT(!fd_isused(fdp, fd)); 220 } else if (dt->dt_ff[fd]->ff_file != NULL) { 221 KASSERT(fd_isused(fdp, fd)); 222 } 223 } 224 #endif 225 } 226 227 static int 228 fd_next_zero(filedesc_t *fdp, uint32_t *bitmap, int want, u_int bits) 229 { 230 int i, off, maxoff; 231 uint32_t sub; 232 233 KASSERT(mutex_owned(&fdp->fd_lock)); 234 235 fd_checkmaps(fdp); 236 237 if (want > bits) 238 return -1; 239 240 off = want >> NDENTRYSHIFT; 241 i = want & NDENTRYMASK; 242 if (i) { 243 sub = bitmap[off] | ((u_int)~0 >> (NDENTRIES - i)); 244 if (sub != ~0) 245 goto found; 246 off++; 247 } 248 249 maxoff = NDLOSLOTS(bits); 250 while (off < maxoff) { 251 if ((sub = bitmap[off]) != ~0) 252 goto found; 253 off++; 254 } 255 256 return -1; 257 258 found: 259 return (off << NDENTRYSHIFT) + ffs(~sub) - 1; 260 } 261 262 static int 263 fd_last_set(filedesc_t *fd, int last) 264 { 265 int off, i; 266 fdfile_t **ff = fd->fd_dt->dt_ff; 267 uint32_t *bitmap = fd->fd_lomap; 268 269 KASSERT(mutex_owned(&fd->fd_lock)); 270 271 fd_checkmaps(fd); 272 273 off = (last - 1) >> NDENTRYSHIFT; 274 275 while (off >= 0 && !bitmap[off]) 276 off--; 277 278 if (off < 0) 279 return -1; 280 281 i = ((off + 1) << NDENTRYSHIFT) - 1; 282 if (i >= last) 283 i = last - 1; 284 285 /* XXX should use bitmap */ 286 while (i > 0 && (ff[i] == NULL || !ff[i]->ff_allocated)) 287 i--; 288 289 return i; 290 } 291 292 static inline void 293 fd_used(filedesc_t *fdp, unsigned fd) 294 { 295 u_int off = fd >> NDENTRYSHIFT; 296 fdfile_t *ff; 297 298 ff = fdp->fd_dt->dt_ff[fd]; 299 300 KASSERT(mutex_owned(&fdp->fd_lock)); 301 KASSERT((fdp->fd_lomap[off] & (1U << (fd & NDENTRYMASK))) == 0); 302 KASSERT(ff != NULL); 303 KASSERT(ff->ff_file == NULL); 304 KASSERT(!ff->ff_allocated); 305 306 ff->ff_allocated = true; 307 fdp->fd_lomap[off] |= 1U << (fd & NDENTRYMASK); 308 if (__predict_false(fdp->fd_lomap[off] == ~0)) { 309 KASSERT((fdp->fd_himap[off >> NDENTRYSHIFT] & 310 (1U << (off & NDENTRYMASK))) == 0); 311 fdp->fd_himap[off >> NDENTRYSHIFT] |= 1U << (off & NDENTRYMASK); 312 } 313 314 if ((int)fd > fdp->fd_lastfile) { 315 fdp->fd_lastfile = fd; 316 } 317 318 fd_checkmaps(fdp); 319 } 320 321 static inline void 322 fd_unused(filedesc_t *fdp, unsigned fd) 323 { 324 u_int off = fd >> NDENTRYSHIFT; 325 fdfile_t *ff; 326 327 ff = fdp->fd_dt->dt_ff[fd]; 328 329 KASSERT(mutex_owned(&fdp->fd_lock)); 330 KASSERT(ff != NULL); 331 KASSERT(ff->ff_file == NULL); 332 KASSERT(ff->ff_allocated); 333 334 if (fd < fdp->fd_freefile) { 335 fdp->fd_freefile = fd; 336 } 337 338 if (fdp->fd_lomap[off] == ~0) { 339 KASSERT((fdp->fd_himap[off >> NDENTRYSHIFT] & 340 (1U << (off & NDENTRYMASK))) != 0); 341 fdp->fd_himap[off >> NDENTRYSHIFT] &= 342 ~(1U << (off & NDENTRYMASK)); 343 } 344 KASSERT((fdp->fd_lomap[off] & (1U << (fd & NDENTRYMASK))) != 0); 345 fdp->fd_lomap[off] &= ~(1U << (fd & NDENTRYMASK)); 346 ff->ff_allocated = false; 347 348 KASSERT(fd <= fdp->fd_lastfile); 349 if (fd == fdp->fd_lastfile) { 350 fdp->fd_lastfile = fd_last_set(fdp, fd); 351 } 352 fd_checkmaps(fdp); 353 } 354 355 /* 356 * Look up the file structure corresponding to a file descriptor 357 * and return the file, holding a reference on the descriptor. 358 */ 359 file_t * 360 fd_getfile(unsigned fd) 361 { 362 filedesc_t *fdp; 363 fdfile_t *ff; 364 file_t *fp; 365 fdtab_t *dt; 366 367 /* 368 * Look up the fdfile structure representing this descriptor. 369 * We are doing this unlocked. See fd_tryexpand(). 370 */ 371 fdp = curlwp->l_fd; 372 dt = atomic_load_consume(&fdp->fd_dt); 373 if (__predict_false(fd >= dt->dt_nfiles)) { 374 return NULL; 375 } 376 ff = dt->dt_ff[fd]; 377 KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]); 378 if (__predict_false(ff == NULL)) { 379 return NULL; 380 } 381 382 /* Now get a reference to the descriptor. */ 383 if (fdp->fd_refcnt == 1) { 384 /* 385 * Single threaded: don't need to worry about concurrent 386 * access (other than earlier calls to kqueue, which may 387 * hold a reference to the descriptor). 388 */ 389 ff->ff_refcnt++; 390 } else { 391 /* 392 * Multi threaded: issue a memory barrier to ensure that we 393 * acquire the file pointer _after_ adding a reference. If 394 * no memory barrier, we could fetch a stale pointer. 395 * 396 * In particular, we must coordinate the following four 397 * memory operations: 398 * 399 * A. fd_close store ff->ff_file = NULL 400 * B. fd_close refcnt = atomic_dec_uint_nv(&ff->ff_refcnt) 401 * C. fd_getfile atomic_inc_uint(&ff->ff_refcnt) 402 * D. fd_getfile load fp = ff->ff_file 403 * 404 * If the order is D;A;B;C: 405 * 406 * 1. D: fp = ff->ff_file 407 * 2. A: ff->ff_file = NULL 408 * 3. B: refcnt = atomic_dec_uint_nv(&ff->ff_refcnt) 409 * 4. C: atomic_inc_uint(&ff->ff_refcnt) 410 * 411 * then fd_close determines that there are no more 412 * references and decides to free fp immediately, at 413 * the same that fd_getfile ends up with an fp that's 414 * about to be freed. *boom* 415 * 416 * By making B a release operation in fd_close, and by 417 * making C an acquire operation in fd_getfile, since 418 * they are atomic operations on the same object, which 419 * has a total modification order, we guarantee either: 420 * 421 * - B happens before C. Then since A is 422 * sequenced before B in fd_close, and C is 423 * sequenced before D in fd_getfile, we 424 * guarantee A happens before D, so fd_getfile 425 * reads a null fp and safely fails. 426 * 427 * - C happens before B. Then fd_getfile may read 428 * null or nonnull, but either way, fd_close 429 * will safely wait for references to drain. 430 */ 431 atomic_inc_uint(&ff->ff_refcnt); 432 membar_acquire(); 433 } 434 435 /* 436 * If the file is not open or is being closed then put the 437 * reference back. 438 */ 439 fp = atomic_load_consume(&ff->ff_file); 440 if (__predict_true(fp != NULL)) { 441 return fp; 442 } 443 fd_putfile(fd); 444 return NULL; 445 } 446 447 /* 448 * Release a reference to a file descriptor acquired with fd_getfile(). 449 */ 450 void 451 fd_putfile(unsigned fd) 452 { 453 filedesc_t *fdp; 454 fdfile_t *ff; 455 u_int u, v; 456 457 fdp = curlwp->l_fd; 458 KASSERT(fd < atomic_load_consume(&fdp->fd_dt)->dt_nfiles); 459 ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd]; 460 461 KASSERT(ff != NULL); 462 KASSERT((ff->ff_refcnt & FR_MASK) > 0); 463 KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]); 464 465 if (fdp->fd_refcnt == 1) { 466 /* 467 * Single threaded: don't need to worry about concurrent 468 * access (other than earlier calls to kqueue, which may 469 * hold a reference to the descriptor). 470 */ 471 if (__predict_false((ff->ff_refcnt & FR_CLOSING) != 0)) { 472 fd_close(fd); 473 return; 474 } 475 ff->ff_refcnt--; 476 return; 477 } 478 479 /* 480 * Ensure that any use of the file is complete and globally 481 * visible before dropping the final reference. If no membar, 482 * the current CPU could still access memory associated with 483 * the file after it has been freed or recycled by another 484 * CPU. 485 */ 486 membar_release(); 487 488 /* 489 * Be optimistic and start out with the assumption that no other 490 * threads are trying to close the descriptor. If the CAS fails, 491 * we lost a race and/or it's being closed. 492 */ 493 for (u = ff->ff_refcnt & FR_MASK;; u = v) { 494 v = atomic_cas_uint(&ff->ff_refcnt, u, u - 1); 495 if (__predict_true(u == v)) { 496 return; 497 } 498 if (__predict_false((v & FR_CLOSING) != 0)) { 499 break; 500 } 501 } 502 503 /* Another thread is waiting to close the file: join it. */ 504 (void)fd_close(fd); 505 } 506 507 /* 508 * Convenience wrapper around fd_getfile() that returns reference 509 * to a vnode. 510 */ 511 int 512 fd_getvnode(unsigned fd, file_t **fpp) 513 { 514 vnode_t *vp; 515 file_t *fp; 516 517 fp = fd_getfile(fd); 518 if (__predict_false(fp == NULL)) { 519 return EBADF; 520 } 521 if (__predict_false(fp->f_type != DTYPE_VNODE)) { 522 fd_putfile(fd); 523 return EINVAL; 524 } 525 vp = fp->f_vnode; 526 if (__predict_false(vp->v_type == VBAD)) { 527 /* XXX Is this case really necessary? */ 528 fd_putfile(fd); 529 return EBADF; 530 } 531 *fpp = fp; 532 return 0; 533 } 534 535 /* 536 * Convenience wrapper around fd_getfile() that returns reference 537 * to a socket. 538 */ 539 int 540 fd_getsock1(unsigned fd, struct socket **sop, file_t **fp) 541 { 542 *fp = fd_getfile(fd); 543 if (__predict_false(*fp == NULL)) { 544 return EBADF; 545 } 546 if (__predict_false((*fp)->f_type != DTYPE_SOCKET)) { 547 fd_putfile(fd); 548 return ENOTSOCK; 549 } 550 *sop = (*fp)->f_socket; 551 return 0; 552 } 553 554 int 555 fd_getsock(unsigned fd, struct socket **sop) 556 { 557 file_t *fp; 558 return fd_getsock1(fd, sop, &fp); 559 } 560 561 /* 562 * Look up the file structure corresponding to a file descriptor 563 * and return it with a reference held on the file, not the 564 * descriptor. 565 * 566 * This is heavyweight and only used when accessing descriptors 567 * from a foreign process. The caller must ensure that `p' does 568 * not exit or fork across this call. 569 * 570 * To release the file (not descriptor) reference, use closef(). 571 */ 572 file_t * 573 fd_getfile2(proc_t *p, unsigned fd) 574 { 575 filedesc_t *fdp; 576 fdfile_t *ff; 577 file_t *fp; 578 fdtab_t *dt; 579 580 fdp = p->p_fd; 581 mutex_enter(&fdp->fd_lock); 582 dt = fdp->fd_dt; 583 if (fd >= dt->dt_nfiles) { 584 mutex_exit(&fdp->fd_lock); 585 return NULL; 586 } 587 if ((ff = dt->dt_ff[fd]) == NULL) { 588 mutex_exit(&fdp->fd_lock); 589 return NULL; 590 } 591 if ((fp = atomic_load_consume(&ff->ff_file)) == NULL) { 592 mutex_exit(&fdp->fd_lock); 593 return NULL; 594 } 595 mutex_enter(&fp->f_lock); 596 fp->f_count++; 597 mutex_exit(&fp->f_lock); 598 mutex_exit(&fdp->fd_lock); 599 600 return fp; 601 } 602 603 /* 604 * Internal form of close. Must be called with a reference to the 605 * descriptor, and will drop the reference. When all descriptor 606 * references are dropped, releases the descriptor slot and a single 607 * reference to the file structure. 608 */ 609 int 610 fd_close(unsigned fd) 611 { 612 struct flock lf; 613 filedesc_t *fdp; 614 fdfile_t *ff; 615 file_t *fp; 616 proc_t *p; 617 lwp_t *l; 618 u_int refcnt; 619 620 l = curlwp; 621 p = l->l_proc; 622 fdp = l->l_fd; 623 ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd]; 624 625 KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]); 626 627 mutex_enter(&fdp->fd_lock); 628 KASSERT((ff->ff_refcnt & FR_MASK) > 0); 629 fp = atomic_load_consume(&ff->ff_file); 630 if (__predict_false(fp == NULL)) { 631 /* 632 * Another user of the file is already closing, and is 633 * waiting for other users of the file to drain. Release 634 * our reference, and wake up the closer. 635 */ 636 membar_release(); 637 atomic_dec_uint(&ff->ff_refcnt); 638 cv_broadcast(&ff->ff_closing); 639 mutex_exit(&fdp->fd_lock); 640 641 /* 642 * An application error, so pretend that the descriptor 643 * was already closed. We can't safely wait for it to 644 * be closed without potentially deadlocking. 645 */ 646 return (EBADF); 647 } 648 KASSERT((ff->ff_refcnt & FR_CLOSING) == 0); 649 650 /* 651 * There may be multiple users of this file within the process. 652 * Notify existing and new users that the file is closing. This 653 * will prevent them from adding additional uses to this file 654 * while we are closing it. 655 */ 656 atomic_store_relaxed(&ff->ff_file, NULL); 657 ff->ff_exclose = false; 658 659 /* 660 * We expect the caller to hold a descriptor reference - drop it. 661 * The reference count may increase beyond zero at this point due 662 * to an erroneous descriptor reference by an application, but 663 * fd_getfile() will notice that the file is being closed and drop 664 * the reference again. 665 */ 666 if (fdp->fd_refcnt == 1) { 667 /* Single threaded. */ 668 refcnt = --(ff->ff_refcnt); 669 } else { 670 /* Multi threaded. */ 671 membar_release(); 672 refcnt = atomic_dec_uint_nv(&ff->ff_refcnt); 673 membar_acquire(); 674 } 675 if (__predict_false(refcnt != 0)) { 676 /* 677 * Wait for other references to drain. This is typically 678 * an application error - the descriptor is being closed 679 * while still in use. 680 * (Or just a threaded application trying to unblock its 681 * thread that sleeps in (say) accept()). 682 */ 683 atomic_or_uint(&ff->ff_refcnt, FR_CLOSING); 684 685 /* 686 * Remove any knotes attached to the file. A knote 687 * attached to the descriptor can hold references on it. 688 */ 689 mutex_exit(&fdp->fd_lock); 690 if (!SLIST_EMPTY(&ff->ff_knlist)) { 691 knote_fdclose(fd); 692 } 693 694 /* 695 * Since the file system code doesn't know which fd 696 * each request came from (think dup()), we have to 697 * ask it to return ERESTART for any long-term blocks. 698 * The re-entry through read/write/etc will detect the 699 * closed fd and return EBAFD. 700 * Blocked partial writes may return a short length. 701 */ 702 (*fp->f_ops->fo_restart)(fp); 703 mutex_enter(&fdp->fd_lock); 704 705 /* 706 * We need to see the count drop to zero at least once, 707 * in order to ensure that all pre-existing references 708 * have been drained. New references past this point are 709 * of no interest. 710 * XXX (dsl) this may need to call fo_restart() after a 711 * timeout to guarantee that all the system calls exit. 712 */ 713 while ((ff->ff_refcnt & FR_MASK) != 0) { 714 cv_wait(&ff->ff_closing, &fdp->fd_lock); 715 } 716 atomic_and_uint(&ff->ff_refcnt, ~FR_CLOSING); 717 } else { 718 /* If no references, there must be no knotes. */ 719 KASSERT(SLIST_EMPTY(&ff->ff_knlist)); 720 } 721 722 /* 723 * POSIX record locking dictates that any close releases ALL 724 * locks owned by this process. This is handled by setting 725 * a flag in the unlock to free ONLY locks obeying POSIX 726 * semantics, and not to free BSD-style file locks. 727 * If the descriptor was in a message, POSIX-style locks 728 * aren't passed with the descriptor. 729 */ 730 if (__predict_false((p->p_flag & PK_ADVLOCK) != 0) && 731 fp->f_ops->fo_advlock != NULL) { 732 lf.l_whence = SEEK_SET; 733 lf.l_start = 0; 734 lf.l_len = 0; 735 lf.l_type = F_UNLCK; 736 mutex_exit(&fdp->fd_lock); 737 (void)(*fp->f_ops->fo_advlock)(fp, p, F_UNLCK, &lf, F_POSIX); 738 mutex_enter(&fdp->fd_lock); 739 } 740 741 /* Free descriptor slot. */ 742 fd_unused(fdp, fd); 743 mutex_exit(&fdp->fd_lock); 744 745 /* Now drop reference to the file itself. */ 746 return closef(fp); 747 } 748 749 /* 750 * Duplicate a file descriptor. 751 */ 752 int 753 fd_dup(file_t *fp, int minfd, int *newp, bool exclose) 754 { 755 proc_t *p = curproc; 756 fdtab_t *dt; 757 int error; 758 759 while ((error = fd_alloc(p, minfd, newp)) != 0) { 760 if (error != ENOSPC) { 761 return error; 762 } 763 fd_tryexpand(p); 764 } 765 766 dt = atomic_load_consume(&curlwp->l_fd->fd_dt); 767 dt->dt_ff[*newp]->ff_exclose = exclose; 768 fd_affix(p, fp, *newp); 769 return 0; 770 } 771 772 /* 773 * dup2 operation. 774 */ 775 int 776 fd_dup2(file_t *fp, unsigned newfd, int flags) 777 { 778 filedesc_t *fdp = curlwp->l_fd; 779 fdfile_t *ff; 780 fdtab_t *dt; 781 782 if (flags & ~(O_CLOEXEC|O_NONBLOCK|O_NOSIGPIPE)) 783 return EINVAL; 784 /* 785 * Ensure there are enough slots in the descriptor table, 786 * and allocate an fdfile_t up front in case we need it. 787 */ 788 while (newfd >= atomic_load_consume(&fdp->fd_dt)->dt_nfiles) { 789 fd_tryexpand(curproc); 790 } 791 ff = pool_cache_get(fdfile_cache, PR_WAITOK); 792 793 /* 794 * If there is already a file open, close it. If the file is 795 * half open, wait for it to be constructed before closing it. 796 * XXX Potential for deadlock here? 797 */ 798 mutex_enter(&fdp->fd_lock); 799 while (fd_isused(fdp, newfd)) { 800 mutex_exit(&fdp->fd_lock); 801 if (fd_getfile(newfd) != NULL) { 802 (void)fd_close(newfd); 803 } else { 804 /* 805 * Crummy, but unlikely to happen. 806 * Can occur if we interrupt another 807 * thread while it is opening a file. 808 */ 809 kpause("dup2", false, 1, NULL); 810 } 811 mutex_enter(&fdp->fd_lock); 812 } 813 dt = fdp->fd_dt; 814 if (dt->dt_ff[newfd] == NULL) { 815 KASSERT(newfd >= NDFDFILE); 816 dt->dt_ff[newfd] = ff; 817 ff = NULL; 818 } 819 fd_used(fdp, newfd); 820 mutex_exit(&fdp->fd_lock); 821 822 dt->dt_ff[newfd]->ff_exclose = (flags & O_CLOEXEC) != 0; 823 fp->f_flag |= flags & (FNONBLOCK|FNOSIGPIPE); 824 /* Slot is now allocated. Insert copy of the file. */ 825 fd_affix(curproc, fp, newfd); 826 if (ff != NULL) { 827 pool_cache_put(fdfile_cache, ff); 828 } 829 return 0; 830 } 831 832 /* 833 * Drop reference to a file structure. 834 */ 835 int 836 closef(file_t *fp) 837 { 838 struct flock lf; 839 int error; 840 841 /* 842 * Drop reference. If referenced elsewhere it's still open 843 * and we have nothing more to do. 844 */ 845 mutex_enter(&fp->f_lock); 846 KASSERT(fp->f_count > 0); 847 if (--fp->f_count > 0) { 848 mutex_exit(&fp->f_lock); 849 return 0; 850 } 851 KASSERT(fp->f_count == 0); 852 mutex_exit(&fp->f_lock); 853 854 /* We held the last reference - release locks, close and free. */ 855 if (fp->f_ops->fo_advlock == NULL) { 856 KASSERT((fp->f_flag & FHASLOCK) == 0); 857 } else if (fp->f_flag & FHASLOCK) { 858 lf.l_whence = SEEK_SET; 859 lf.l_start = 0; 860 lf.l_len = 0; 861 lf.l_type = F_UNLCK; 862 (void)(*fp->f_ops->fo_advlock)(fp, fp, F_UNLCK, &lf, F_FLOCK); 863 } 864 if (fp->f_ops != NULL) { 865 error = (*fp->f_ops->fo_close)(fp); 866 } else { 867 error = 0; 868 } 869 KASSERT(fp->f_count == 0); 870 KASSERT(fp->f_cred != NULL); 871 pool_cache_put(file_cache, fp); 872 873 return error; 874 } 875 876 /* 877 * Allocate a file descriptor for the process. 878 */ 879 int 880 fd_alloc(proc_t *p, int want, int *result) 881 { 882 filedesc_t *fdp = p->p_fd; 883 int i, lim, last, error, hi; 884 u_int off; 885 fdtab_t *dt; 886 887 KASSERT(p == curproc || p == &proc0); 888 889 /* 890 * Search for a free descriptor starting at the higher 891 * of want or fd_freefile. 892 */ 893 mutex_enter(&fdp->fd_lock); 894 fd_checkmaps(fdp); 895 dt = fdp->fd_dt; 896 KASSERT(dt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]); 897 lim = uimin((int)p->p_rlimit[RLIMIT_NOFILE].rlim_cur, maxfiles); 898 last = uimin(dt->dt_nfiles, lim); 899 for (;;) { 900 if ((i = want) < fdp->fd_freefile) 901 i = fdp->fd_freefile; 902 off = i >> NDENTRYSHIFT; 903 hi = fd_next_zero(fdp, fdp->fd_himap, off, 904 (last + NDENTRIES - 1) >> NDENTRYSHIFT); 905 if (hi == -1) 906 break; 907 i = fd_next_zero(fdp, &fdp->fd_lomap[hi], 908 hi > off ? 0 : i & NDENTRYMASK, NDENTRIES); 909 if (i == -1) { 910 /* 911 * Free file descriptor in this block was 912 * below want, try again with higher want. 913 */ 914 want = (hi + 1) << NDENTRYSHIFT; 915 continue; 916 } 917 i += (hi << NDENTRYSHIFT); 918 if (i >= last) { 919 break; 920 } 921 if (dt->dt_ff[i] == NULL) { 922 KASSERT(i >= NDFDFILE); 923 dt->dt_ff[i] = pool_cache_get(fdfile_cache, PR_WAITOK); 924 } 925 KASSERT(dt->dt_ff[i]->ff_file == NULL); 926 fd_used(fdp, i); 927 if (want <= fdp->fd_freefile) { 928 fdp->fd_freefile = i; 929 } 930 *result = i; 931 KASSERT(i >= NDFDFILE || 932 dt->dt_ff[i] == (fdfile_t *)fdp->fd_dfdfile[i]); 933 fd_checkmaps(fdp); 934 mutex_exit(&fdp->fd_lock); 935 return 0; 936 } 937 938 /* No space in current array. Let the caller expand and retry. */ 939 error = (dt->dt_nfiles >= lim) ? EMFILE : ENOSPC; 940 mutex_exit(&fdp->fd_lock); 941 return error; 942 } 943 944 /* 945 * Allocate memory for a descriptor table. 946 */ 947 static fdtab_t * 948 fd_dtab_alloc(int n) 949 { 950 fdtab_t *dt; 951 size_t sz; 952 953 KASSERT(n > NDFILE); 954 955 sz = sizeof(*dt) + (n - NDFILE) * sizeof(dt->dt_ff[0]); 956 dt = kmem_alloc(sz, KM_SLEEP); 957 #ifdef DIAGNOSTIC 958 memset(dt, 0xff, sz); 959 #endif 960 dt->dt_nfiles = n; 961 dt->dt_link = NULL; 962 return dt; 963 } 964 965 /* 966 * Free a descriptor table, and all tables linked for deferred free. 967 */ 968 static void 969 fd_dtab_free(fdtab_t *dt) 970 { 971 fdtab_t *next; 972 size_t sz; 973 974 do { 975 next = dt->dt_link; 976 KASSERT(dt->dt_nfiles > NDFILE); 977 sz = sizeof(*dt) + 978 (dt->dt_nfiles - NDFILE) * sizeof(dt->dt_ff[0]); 979 #ifdef DIAGNOSTIC 980 memset(dt, 0xff, sz); 981 #endif 982 kmem_free(dt, sz); 983 dt = next; 984 } while (dt != NULL); 985 } 986 987 /* 988 * Allocate descriptor bitmap. 989 */ 990 static void 991 fd_map_alloc(int n, uint32_t **lo, uint32_t **hi) 992 { 993 uint8_t *ptr; 994 size_t szlo, szhi; 995 996 KASSERT(n > NDENTRIES); 997 998 szlo = NDLOSLOTS(n) * sizeof(uint32_t); 999 szhi = NDHISLOTS(n) * sizeof(uint32_t); 1000 ptr = kmem_alloc(szlo + szhi, KM_SLEEP); 1001 *lo = (uint32_t *)ptr; 1002 *hi = (uint32_t *)(ptr + szlo); 1003 } 1004 1005 /* 1006 * Free descriptor bitmap. 1007 */ 1008 static void 1009 fd_map_free(int n, uint32_t *lo, uint32_t *hi) 1010 { 1011 size_t szlo, szhi; 1012 1013 KASSERT(n > NDENTRIES); 1014 1015 szlo = NDLOSLOTS(n) * sizeof(uint32_t); 1016 szhi = NDHISLOTS(n) * sizeof(uint32_t); 1017 KASSERT(hi == (uint32_t *)((uint8_t *)lo + szlo)); 1018 kmem_free(lo, szlo + szhi); 1019 } 1020 1021 /* 1022 * Expand a process' descriptor table. 1023 */ 1024 void 1025 fd_tryexpand(proc_t *p) 1026 { 1027 filedesc_t *fdp; 1028 int i, numfiles, oldnfiles; 1029 fdtab_t *newdt, *dt; 1030 uint32_t *newhimap, *newlomap; 1031 1032 KASSERT(p == curproc || p == &proc0); 1033 1034 fdp = p->p_fd; 1035 newhimap = NULL; 1036 newlomap = NULL; 1037 oldnfiles = atomic_load_consume(&fdp->fd_dt)->dt_nfiles; 1038 1039 if (oldnfiles < NDEXTENT) 1040 numfiles = NDEXTENT; 1041 else 1042 numfiles = 2 * oldnfiles; 1043 1044 newdt = fd_dtab_alloc(numfiles); 1045 if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) { 1046 fd_map_alloc(numfiles, &newlomap, &newhimap); 1047 } 1048 1049 mutex_enter(&fdp->fd_lock); 1050 dt = fdp->fd_dt; 1051 KASSERT(dt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]); 1052 if (dt->dt_nfiles != oldnfiles) { 1053 /* fdp changed; caller must retry */ 1054 mutex_exit(&fdp->fd_lock); 1055 fd_dtab_free(newdt); 1056 if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) { 1057 fd_map_free(numfiles, newlomap, newhimap); 1058 } 1059 return; 1060 } 1061 1062 /* Copy the existing descriptor table and zero the new portion. */ 1063 i = sizeof(fdfile_t *) * oldnfiles; 1064 memcpy(newdt->dt_ff, dt->dt_ff, i); 1065 memset((uint8_t *)newdt->dt_ff + i, 0, 1066 numfiles * sizeof(fdfile_t *) - i); 1067 1068 /* 1069 * Link old descriptor array into list to be discarded. We defer 1070 * freeing until the last reference to the descriptor table goes 1071 * away (usually process exit). This allows us to do lockless 1072 * lookups in fd_getfile(). 1073 */ 1074 if (oldnfiles > NDFILE) { 1075 if (fdp->fd_refcnt > 1) { 1076 newdt->dt_link = dt; 1077 } else { 1078 fd_dtab_free(dt); 1079 } 1080 } 1081 1082 if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) { 1083 i = NDHISLOTS(oldnfiles) * sizeof(uint32_t); 1084 memcpy(newhimap, fdp->fd_himap, i); 1085 memset((uint8_t *)newhimap + i, 0, 1086 NDHISLOTS(numfiles) * sizeof(uint32_t) - i); 1087 1088 i = NDLOSLOTS(oldnfiles) * sizeof(uint32_t); 1089 memcpy(newlomap, fdp->fd_lomap, i); 1090 memset((uint8_t *)newlomap + i, 0, 1091 NDLOSLOTS(numfiles) * sizeof(uint32_t) - i); 1092 1093 if (NDHISLOTS(oldnfiles) > NDHISLOTS(NDFILE)) { 1094 fd_map_free(oldnfiles, fdp->fd_lomap, fdp->fd_himap); 1095 } 1096 fdp->fd_himap = newhimap; 1097 fdp->fd_lomap = newlomap; 1098 } 1099 1100 /* 1101 * All other modifications must become globally visible before 1102 * the change to fd_dt. See fd_getfile(). 1103 */ 1104 atomic_store_release(&fdp->fd_dt, newdt); 1105 KASSERT(newdt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]); 1106 fd_checkmaps(fdp); 1107 mutex_exit(&fdp->fd_lock); 1108 } 1109 1110 /* 1111 * Create a new open file structure and allocate a file descriptor 1112 * for the current process. 1113 */ 1114 int 1115 fd_allocfile(file_t **resultfp, int *resultfd) 1116 { 1117 proc_t *p = curproc; 1118 kauth_cred_t cred; 1119 file_t *fp; 1120 int error; 1121 1122 while ((error = fd_alloc(p, 0, resultfd)) != 0) { 1123 if (error != ENOSPC) { 1124 return error; 1125 } 1126 fd_tryexpand(p); 1127 } 1128 1129 fp = pool_cache_get(file_cache, PR_WAITOK); 1130 if (fp == NULL) { 1131 fd_abort(p, NULL, *resultfd); 1132 return ENFILE; 1133 } 1134 KASSERT(fp->f_count == 0); 1135 KASSERT(fp->f_msgcount == 0); 1136 KASSERT(fp->f_unpcount == 0); 1137 1138 /* Replace cached credentials if not what we need. */ 1139 cred = curlwp->l_cred; 1140 if (__predict_false(cred != fp->f_cred)) { 1141 kauth_cred_free(fp->f_cred); 1142 kauth_cred_hold(cred); 1143 fp->f_cred = cred; 1144 } 1145 1146 /* 1147 * Don't allow recycled files to be scanned. 1148 * See uipc_usrreq.c. 1149 */ 1150 if (__predict_false((fp->f_flag & FSCAN) != 0)) { 1151 mutex_enter(&fp->f_lock); 1152 atomic_and_uint(&fp->f_flag, ~FSCAN); 1153 mutex_exit(&fp->f_lock); 1154 } 1155 1156 fp->f_advice = 0; 1157 fp->f_offset = 0; 1158 *resultfp = fp; 1159 1160 return 0; 1161 } 1162 1163 /* 1164 * Successful creation of a new descriptor: make visible to the process. 1165 */ 1166 void 1167 fd_affix(proc_t *p, file_t *fp, unsigned fd) 1168 { 1169 fdfile_t *ff; 1170 filedesc_t *fdp; 1171 fdtab_t *dt; 1172 1173 KASSERT(p == curproc || p == &proc0); 1174 1175 /* Add a reference to the file structure. */ 1176 mutex_enter(&fp->f_lock); 1177 fp->f_count++; 1178 mutex_exit(&fp->f_lock); 1179 1180 /* 1181 * Insert the new file into the descriptor slot. 1182 */ 1183 fdp = p->p_fd; 1184 dt = atomic_load_consume(&fdp->fd_dt); 1185 ff = dt->dt_ff[fd]; 1186 1187 KASSERT(ff != NULL); 1188 KASSERT(ff->ff_file == NULL); 1189 KASSERT(ff->ff_allocated); 1190 KASSERT(fd_isused(fdp, fd)); 1191 KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]); 1192 1193 /* No need to lock in order to make file initially visible. */ 1194 atomic_store_release(&ff->ff_file, fp); 1195 } 1196 1197 /* 1198 * Abort creation of a new descriptor: free descriptor slot and file. 1199 */ 1200 void 1201 fd_abort(proc_t *p, file_t *fp, unsigned fd) 1202 { 1203 filedesc_t *fdp; 1204 fdfile_t *ff; 1205 1206 KASSERT(p == curproc || p == &proc0); 1207 1208 fdp = p->p_fd; 1209 ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd]; 1210 ff->ff_exclose = false; 1211 1212 KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]); 1213 1214 mutex_enter(&fdp->fd_lock); 1215 KASSERT(fd_isused(fdp, fd)); 1216 fd_unused(fdp, fd); 1217 mutex_exit(&fdp->fd_lock); 1218 1219 if (fp != NULL) { 1220 KASSERT(fp->f_count == 0); 1221 KASSERT(fp->f_cred != NULL); 1222 pool_cache_put(file_cache, fp); 1223 } 1224 } 1225 1226 static int 1227 file_ctor(void *arg, void *obj, int flags) 1228 { 1229 /* 1230 * It's easy to exhaust the open file limit on a system with many 1231 * CPUs due to caching. Allow a bit of leeway to reduce the element 1232 * of surprise. 1233 */ 1234 u_int slop = PCG_NOBJECTS_NORMAL * (ncpu - 1); 1235 file_t *fp = obj; 1236 1237 memset(fp, 0, sizeof(*fp)); 1238 1239 mutex_enter(&filelist_lock); 1240 if (__predict_false(nfiles >= slop + maxfiles)) { 1241 mutex_exit(&filelist_lock); 1242 tablefull("file", "increase kern.maxfiles or MAXFILES"); 1243 return ENFILE; 1244 } 1245 nfiles++; 1246 LIST_INSERT_HEAD(&filehead, fp, f_list); 1247 mutex_init(&fp->f_lock, MUTEX_DEFAULT, IPL_NONE); 1248 fp->f_cred = curlwp->l_cred; 1249 kauth_cred_hold(fp->f_cred); 1250 mutex_exit(&filelist_lock); 1251 1252 return 0; 1253 } 1254 1255 static void 1256 file_dtor(void *arg, void *obj) 1257 { 1258 file_t *fp = obj; 1259 1260 mutex_enter(&filelist_lock); 1261 nfiles--; 1262 LIST_REMOVE(fp, f_list); 1263 mutex_exit(&filelist_lock); 1264 1265 KASSERT(fp->f_count == 0); 1266 kauth_cred_free(fp->f_cred); 1267 mutex_destroy(&fp->f_lock); 1268 } 1269 1270 static int 1271 fdfile_ctor(void *arg, void *obj, int flags) 1272 { 1273 fdfile_t *ff = obj; 1274 1275 memset(ff, 0, sizeof(*ff)); 1276 cv_init(&ff->ff_closing, "fdclose"); 1277 1278 return 0; 1279 } 1280 1281 static void 1282 fdfile_dtor(void *arg, void *obj) 1283 { 1284 fdfile_t *ff = obj; 1285 1286 cv_destroy(&ff->ff_closing); 1287 } 1288 1289 file_t * 1290 fgetdummy(void) 1291 { 1292 file_t *fp; 1293 1294 fp = kmem_zalloc(sizeof(*fp), KM_SLEEP); 1295 mutex_init(&fp->f_lock, MUTEX_DEFAULT, IPL_NONE); 1296 return fp; 1297 } 1298 1299 void 1300 fputdummy(file_t *fp) 1301 { 1302 1303 mutex_destroy(&fp->f_lock); 1304 kmem_free(fp, sizeof(*fp)); 1305 } 1306 1307 /* 1308 * Create an initial filedesc structure. 1309 */ 1310 filedesc_t * 1311 fd_init(filedesc_t *fdp) 1312 { 1313 #ifdef DIAGNOSTIC 1314 unsigned fd; 1315 #endif 1316 1317 if (__predict_true(fdp == NULL)) { 1318 fdp = pool_cache_get(filedesc_cache, PR_WAITOK); 1319 } else { 1320 KASSERT(fdp == &filedesc0); 1321 filedesc_ctor(NULL, fdp, PR_WAITOK); 1322 } 1323 1324 #ifdef DIAGNOSTIC 1325 KASSERT(fdp->fd_lastfile == -1); 1326 KASSERT(fdp->fd_lastkqfile == -1); 1327 KASSERT(fdp->fd_knhash == NULL); 1328 KASSERT(fdp->fd_freefile == 0); 1329 KASSERT(fdp->fd_exclose == false); 1330 KASSERT(fdp->fd_dt == &fdp->fd_dtbuiltin); 1331 KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE); 1332 for (fd = 0; fd < NDFDFILE; fd++) { 1333 KASSERT(fdp->fd_dtbuiltin.dt_ff[fd] == 1334 (fdfile_t *)fdp->fd_dfdfile[fd]); 1335 } 1336 for (fd = NDFDFILE; fd < NDFILE; fd++) { 1337 KASSERT(fdp->fd_dtbuiltin.dt_ff[fd] == NULL); 1338 } 1339 KASSERT(fdp->fd_himap == fdp->fd_dhimap); 1340 KASSERT(fdp->fd_lomap == fdp->fd_dlomap); 1341 #endif /* DIAGNOSTIC */ 1342 1343 fdp->fd_refcnt = 1; 1344 fd_checkmaps(fdp); 1345 1346 return fdp; 1347 } 1348 1349 /* 1350 * Initialize a file descriptor table. 1351 */ 1352 static int 1353 filedesc_ctor(void *arg, void *obj, int flag) 1354 { 1355 filedesc_t *fdp = obj; 1356 fdfile_t **ffp; 1357 int i; 1358 1359 memset(fdp, 0, sizeof(*fdp)); 1360 mutex_init(&fdp->fd_lock, MUTEX_DEFAULT, IPL_NONE); 1361 fdp->fd_lastfile = -1; 1362 fdp->fd_lastkqfile = -1; 1363 fdp->fd_dt = &fdp->fd_dtbuiltin; 1364 fdp->fd_dtbuiltin.dt_nfiles = NDFILE; 1365 fdp->fd_himap = fdp->fd_dhimap; 1366 fdp->fd_lomap = fdp->fd_dlomap; 1367 1368 CTASSERT(sizeof(fdp->fd_dfdfile[0]) >= sizeof(fdfile_t)); 1369 for (i = 0, ffp = fdp->fd_dt->dt_ff; i < NDFDFILE; i++, ffp++) { 1370 *ffp = (fdfile_t *)fdp->fd_dfdfile[i]; 1371 (void)fdfile_ctor(NULL, fdp->fd_dfdfile[i], PR_WAITOK); 1372 } 1373 1374 return 0; 1375 } 1376 1377 static void 1378 filedesc_dtor(void *arg, void *obj) 1379 { 1380 filedesc_t *fdp = obj; 1381 int i; 1382 1383 for (i = 0; i < NDFDFILE; i++) { 1384 fdfile_dtor(NULL, fdp->fd_dfdfile[i]); 1385 } 1386 1387 mutex_destroy(&fdp->fd_lock); 1388 } 1389 1390 /* 1391 * Make p share curproc's filedesc structure. 1392 */ 1393 void 1394 fd_share(struct proc *p) 1395 { 1396 filedesc_t *fdp; 1397 1398 fdp = curlwp->l_fd; 1399 p->p_fd = fdp; 1400 atomic_inc_uint(&fdp->fd_refcnt); 1401 } 1402 1403 /* 1404 * Acquire a hold on a filedesc structure. 1405 */ 1406 void 1407 fd_hold(lwp_t *l) 1408 { 1409 filedesc_t *fdp = l->l_fd; 1410 1411 atomic_inc_uint(&fdp->fd_refcnt); 1412 } 1413 1414 /* 1415 * Copy a filedesc structure. 1416 */ 1417 filedesc_t * 1418 fd_copy(void) 1419 { 1420 filedesc_t *newfdp, *fdp; 1421 fdfile_t *ff, **ffp, **nffp, *ff2; 1422 int i, j, numfiles, lastfile, newlast; 1423 file_t *fp; 1424 fdtab_t *newdt; 1425 1426 fdp = curproc->p_fd; 1427 newfdp = pool_cache_get(filedesc_cache, PR_WAITOK); 1428 newfdp->fd_refcnt = 1; 1429 1430 #ifdef DIAGNOSTIC 1431 KASSERT(newfdp->fd_lastfile == -1); 1432 KASSERT(newfdp->fd_lastkqfile == -1); 1433 KASSERT(newfdp->fd_knhash == NULL); 1434 KASSERT(newfdp->fd_freefile == 0); 1435 KASSERT(newfdp->fd_exclose == false); 1436 KASSERT(newfdp->fd_dt == &newfdp->fd_dtbuiltin); 1437 KASSERT(newfdp->fd_dtbuiltin.dt_nfiles == NDFILE); 1438 for (i = 0; i < NDFDFILE; i++) { 1439 KASSERT(newfdp->fd_dtbuiltin.dt_ff[i] == 1440 (fdfile_t *)&newfdp->fd_dfdfile[i]); 1441 } 1442 for (i = NDFDFILE; i < NDFILE; i++) { 1443 KASSERT(newfdp->fd_dtbuiltin.dt_ff[i] == NULL); 1444 } 1445 #endif /* DIAGNOSTIC */ 1446 1447 mutex_enter(&fdp->fd_lock); 1448 fd_checkmaps(fdp); 1449 numfiles = fdp->fd_dt->dt_nfiles; 1450 lastfile = fdp->fd_lastfile; 1451 1452 /* 1453 * If the number of open files fits in the internal arrays 1454 * of the open file structure, use them, otherwise allocate 1455 * additional memory for the number of descriptors currently 1456 * in use. 1457 */ 1458 if (lastfile < NDFILE) { 1459 i = NDFILE; 1460 newdt = newfdp->fd_dt; 1461 KASSERT(newfdp->fd_dt == &newfdp->fd_dtbuiltin); 1462 } else { 1463 /* 1464 * Compute the smallest multiple of NDEXTENT needed 1465 * for the file descriptors currently in use, 1466 * allowing the table to shrink. 1467 */ 1468 i = numfiles; 1469 while (i >= 2 * NDEXTENT && i > lastfile * 2) { 1470 i /= 2; 1471 } 1472 KASSERT(i > NDFILE); 1473 newdt = fd_dtab_alloc(i); 1474 newfdp->fd_dt = newdt; 1475 memcpy(newdt->dt_ff, newfdp->fd_dtbuiltin.dt_ff, 1476 NDFDFILE * sizeof(fdfile_t **)); 1477 memset(newdt->dt_ff + NDFDFILE, 0, 1478 (i - NDFDFILE) * sizeof(fdfile_t **)); 1479 } 1480 if (NDHISLOTS(i) <= NDHISLOTS(NDFILE)) { 1481 newfdp->fd_himap = newfdp->fd_dhimap; 1482 newfdp->fd_lomap = newfdp->fd_dlomap; 1483 } else { 1484 fd_map_alloc(i, &newfdp->fd_lomap, &newfdp->fd_himap); 1485 KASSERT(i >= NDENTRIES * NDENTRIES); 1486 memset(newfdp->fd_himap, 0, NDHISLOTS(i)*sizeof(uint32_t)); 1487 memset(newfdp->fd_lomap, 0, NDLOSLOTS(i)*sizeof(uint32_t)); 1488 } 1489 newfdp->fd_freefile = fdp->fd_freefile; 1490 newfdp->fd_exclose = fdp->fd_exclose; 1491 1492 ffp = fdp->fd_dt->dt_ff; 1493 nffp = newdt->dt_ff; 1494 newlast = -1; 1495 for (i = 0; i <= lastfile; i++, ffp++, nffp++) { 1496 KASSERT(i >= NDFDFILE || 1497 *nffp == (fdfile_t *)newfdp->fd_dfdfile[i]); 1498 ff = *ffp; 1499 if (ff == NULL || 1500 (fp = atomic_load_consume(&ff->ff_file)) == NULL) { 1501 /* Descriptor unused, or descriptor half open. */ 1502 KASSERT(!fd_isused(newfdp, i)); 1503 continue; 1504 } 1505 if (__predict_false(fp->f_type == DTYPE_KQUEUE)) { 1506 /* kqueue descriptors cannot be copied. */ 1507 if (i < newfdp->fd_freefile) { 1508 newfdp->fd_freefile = i; 1509 } 1510 continue; 1511 } 1512 /* It's active: add a reference to the file. */ 1513 mutex_enter(&fp->f_lock); 1514 fp->f_count++; 1515 mutex_exit(&fp->f_lock); 1516 1517 /* Allocate an fdfile_t to represent it. */ 1518 if (i >= NDFDFILE) { 1519 ff2 = pool_cache_get(fdfile_cache, PR_WAITOK); 1520 *nffp = ff2; 1521 } else { 1522 ff2 = newdt->dt_ff[i]; 1523 } 1524 ff2->ff_file = fp; 1525 ff2->ff_exclose = ff->ff_exclose; 1526 ff2->ff_allocated = true; 1527 1528 /* Fix up bitmaps. */ 1529 j = i >> NDENTRYSHIFT; 1530 KASSERT((newfdp->fd_lomap[j] & (1U << (i & NDENTRYMASK))) == 0); 1531 newfdp->fd_lomap[j] |= 1U << (i & NDENTRYMASK); 1532 if (__predict_false(newfdp->fd_lomap[j] == ~0)) { 1533 KASSERT((newfdp->fd_himap[j >> NDENTRYSHIFT] & 1534 (1U << (j & NDENTRYMASK))) == 0); 1535 newfdp->fd_himap[j >> NDENTRYSHIFT] |= 1536 1U << (j & NDENTRYMASK); 1537 } 1538 newlast = i; 1539 } 1540 KASSERT(newdt->dt_ff[0] == (fdfile_t *)newfdp->fd_dfdfile[0]); 1541 newfdp->fd_lastfile = newlast; 1542 fd_checkmaps(newfdp); 1543 mutex_exit(&fdp->fd_lock); 1544 1545 return newfdp; 1546 } 1547 1548 /* 1549 * Release a filedesc structure. 1550 */ 1551 void 1552 fd_free(void) 1553 { 1554 fdfile_t *ff; 1555 file_t *fp; 1556 int fd, nf; 1557 fdtab_t *dt; 1558 lwp_t * const l = curlwp; 1559 filedesc_t * const fdp = l->l_fd; 1560 const bool noadvlock = (l->l_proc->p_flag & PK_ADVLOCK) == 0; 1561 1562 KASSERT(atomic_load_consume(&fdp->fd_dt)->dt_ff[0] == 1563 (fdfile_t *)fdp->fd_dfdfile[0]); 1564 KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE); 1565 KASSERT(fdp->fd_dtbuiltin.dt_link == NULL); 1566 1567 membar_release(); 1568 if (atomic_dec_uint_nv(&fdp->fd_refcnt) > 0) 1569 return; 1570 membar_acquire(); 1571 1572 /* 1573 * Close any files that the process holds open. 1574 */ 1575 dt = fdp->fd_dt; 1576 fd_checkmaps(fdp); 1577 #ifdef DEBUG 1578 fdp->fd_refcnt = -1; /* see fd_checkmaps */ 1579 #endif 1580 for (fd = 0, nf = dt->dt_nfiles; fd < nf; fd++) { 1581 ff = dt->dt_ff[fd]; 1582 KASSERT(fd >= NDFDFILE || 1583 ff == (fdfile_t *)fdp->fd_dfdfile[fd]); 1584 if (ff == NULL) 1585 continue; 1586 if ((fp = atomic_load_consume(&ff->ff_file)) != NULL) { 1587 /* 1588 * Must use fd_close() here if there is 1589 * a reference from kqueue or we might have posix 1590 * advisory locks. 1591 */ 1592 if (__predict_true(ff->ff_refcnt == 0) && 1593 (noadvlock || fp->f_type != DTYPE_VNODE)) { 1594 ff->ff_file = NULL; 1595 ff->ff_exclose = false; 1596 ff->ff_allocated = false; 1597 closef(fp); 1598 } else { 1599 ff->ff_refcnt++; 1600 fd_close(fd); 1601 } 1602 } 1603 KASSERT(ff->ff_refcnt == 0); 1604 KASSERT(ff->ff_file == NULL); 1605 KASSERT(!ff->ff_exclose); 1606 KASSERT(!ff->ff_allocated); 1607 if (fd >= NDFDFILE) { 1608 pool_cache_put(fdfile_cache, ff); 1609 dt->dt_ff[fd] = NULL; 1610 } 1611 } 1612 1613 /* 1614 * Clean out the descriptor table for the next user and return 1615 * to the cache. 1616 */ 1617 if (__predict_false(dt != &fdp->fd_dtbuiltin)) { 1618 fd_dtab_free(fdp->fd_dt); 1619 /* Otherwise, done above. */ 1620 memset(&fdp->fd_dtbuiltin.dt_ff[NDFDFILE], 0, 1621 (NDFILE - NDFDFILE) * sizeof(fdp->fd_dtbuiltin.dt_ff[0])); 1622 fdp->fd_dt = &fdp->fd_dtbuiltin; 1623 } 1624 if (__predict_false(NDHISLOTS(nf) > NDHISLOTS(NDFILE))) { 1625 KASSERT(fdp->fd_himap != fdp->fd_dhimap); 1626 KASSERT(fdp->fd_lomap != fdp->fd_dlomap); 1627 fd_map_free(nf, fdp->fd_lomap, fdp->fd_himap); 1628 } 1629 if (__predict_false(fdp->fd_knhash != NULL)) { 1630 hashdone(fdp->fd_knhash, HASH_LIST, fdp->fd_knhashmask); 1631 fdp->fd_knhash = NULL; 1632 fdp->fd_knhashmask = 0; 1633 } else { 1634 KASSERT(fdp->fd_knhashmask == 0); 1635 } 1636 fdp->fd_dt = &fdp->fd_dtbuiltin; 1637 fdp->fd_lastkqfile = -1; 1638 fdp->fd_lastfile = -1; 1639 fdp->fd_freefile = 0; 1640 fdp->fd_exclose = false; 1641 memset(&fdp->fd_startzero, 0, sizeof(*fdp) - 1642 offsetof(filedesc_t, fd_startzero)); 1643 fdp->fd_himap = fdp->fd_dhimap; 1644 fdp->fd_lomap = fdp->fd_dlomap; 1645 KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE); 1646 KASSERT(fdp->fd_dtbuiltin.dt_link == NULL); 1647 KASSERT(fdp->fd_dt == &fdp->fd_dtbuiltin); 1648 #ifdef DEBUG 1649 fdp->fd_refcnt = 0; /* see fd_checkmaps */ 1650 #endif 1651 fd_checkmaps(fdp); 1652 pool_cache_put(filedesc_cache, fdp); 1653 } 1654 1655 /* 1656 * File Descriptor pseudo-device driver (/dev/fd/). 1657 * 1658 * Opening minor device N dup()s the file (if any) connected to file 1659 * descriptor N belonging to the calling process. Note that this driver 1660 * consists of only the ``open()'' routine, because all subsequent 1661 * references to this file will be direct to the other driver. 1662 */ 1663 static int 1664 filedescopen(dev_t dev, int mode, int type, lwp_t *l) 1665 { 1666 1667 /* 1668 * XXX Kludge: set dupfd to contain the value of the 1669 * the file descriptor being sought for duplication. The error 1670 * return ensures that the vnode for this device will be released 1671 * by vn_open. Open will detect this special error and take the 1672 * actions in fd_dupopen below. Other callers of vn_open or VOP_OPEN 1673 * will simply report the error. 1674 */ 1675 l->l_dupfd = minor(dev); /* XXX */ 1676 return EDUPFD; 1677 } 1678 1679 /* 1680 * Duplicate the specified descriptor to a free descriptor. 1681 * 1682 * old is the original fd. 1683 * moveit is true if we should move rather than duplicate. 1684 * flags are the open flags (converted from O_* to F*). 1685 * newp returns the new fd on success. 1686 * 1687 * These two cases are produced by the EDUPFD and EMOVEFD magic 1688 * errnos, but in the interest of removing that regrettable interface, 1689 * vn_open has been changed to intercept them. Now vn_open returns 1690 * either a vnode or a filehandle, and the filehandle is accompanied 1691 * by a boolean that says whether we should dup (moveit == false) or 1692 * move (moveit == true) the fd. 1693 * 1694 * The dup case is used by /dev/stderr, /proc/self/fd, and such. The 1695 * move case is used by cloner devices that allocate a fd of their 1696 * own (a layering violation that should go away eventually) that 1697 * then needs to be put in the place open() expects it. 1698 */ 1699 int 1700 fd_dupopen(int old, bool moveit, int flags, int *newp) 1701 { 1702 filedesc_t *fdp; 1703 fdfile_t *ff; 1704 file_t *fp; 1705 fdtab_t *dt; 1706 int error; 1707 1708 if ((fp = fd_getfile(old)) == NULL) { 1709 return EBADF; 1710 } 1711 fdp = curlwp->l_fd; 1712 dt = atomic_load_consume(&fdp->fd_dt); 1713 ff = dt->dt_ff[old]; 1714 1715 /* 1716 * There are two cases of interest here. 1717 * 1718 * 1. moveit == false (used to be the EDUPFD magic errno): 1719 * simply dup (old) to file descriptor (new) and return. 1720 * 1721 * 2. moveit == true (used to be the EMOVEFD magic errno): 1722 * steal away the file structure from (old) and store it in 1723 * (new). (old) is effectively closed by this operation. 1724 */ 1725 if (moveit == false) { 1726 /* 1727 * Check that the mode the file is being opened for is a 1728 * subset of the mode of the existing descriptor. 1729 */ 1730 if (((flags & (FREAD|FWRITE)) | fp->f_flag) != fp->f_flag) { 1731 error = EACCES; 1732 goto out; 1733 } 1734 1735 /* Copy it. */ 1736 error = fd_dup(fp, 0, newp, ff->ff_exclose); 1737 } else { 1738 /* Copy it. */ 1739 error = fd_dup(fp, 0, newp, ff->ff_exclose); 1740 if (error != 0) { 1741 goto out; 1742 } 1743 1744 /* Steal away the file pointer from 'old'. */ 1745 (void)fd_close(old); 1746 return 0; 1747 } 1748 1749 out: 1750 fd_putfile(old); 1751 return error; 1752 } 1753 1754 /* 1755 * Close open files on exec. 1756 */ 1757 void 1758 fd_closeexec(void) 1759 { 1760 proc_t *p; 1761 filedesc_t *fdp; 1762 fdfile_t *ff; 1763 lwp_t *l; 1764 fdtab_t *dt; 1765 int fd; 1766 1767 l = curlwp; 1768 p = l->l_proc; 1769 fdp = p->p_fd; 1770 1771 if (fdp->fd_refcnt > 1) { 1772 fdp = fd_copy(); 1773 fd_free(); 1774 p->p_fd = fdp; 1775 l->l_fd = fdp; 1776 } 1777 if (!fdp->fd_exclose) { 1778 return; 1779 } 1780 fdp->fd_exclose = false; 1781 dt = atomic_load_consume(&fdp->fd_dt); 1782 1783 for (fd = 0; fd <= fdp->fd_lastfile; fd++) { 1784 if ((ff = dt->dt_ff[fd]) == NULL) { 1785 KASSERT(fd >= NDFDFILE); 1786 continue; 1787 } 1788 KASSERT(fd >= NDFDFILE || 1789 ff == (fdfile_t *)fdp->fd_dfdfile[fd]); 1790 if (ff->ff_file == NULL) 1791 continue; 1792 if (ff->ff_exclose) { 1793 /* 1794 * We need a reference to close the file. 1795 * No other threads can see the fdfile_t at 1796 * this point, so don't bother locking. 1797 */ 1798 KASSERT((ff->ff_refcnt & FR_CLOSING) == 0); 1799 ff->ff_refcnt++; 1800 fd_close(fd); 1801 } 1802 } 1803 } 1804 1805 /* 1806 * Sets descriptor owner. If the owner is a process, 'pgid' 1807 * is set to positive value, process ID. If the owner is process group, 1808 * 'pgid' is set to -pg_id. 1809 */ 1810 int 1811 fsetown(pid_t *pgid, u_long cmd, const void *data) 1812 { 1813 pid_t id = *(const pid_t *)data; 1814 int error; 1815 1816 if (id == INT_MIN) 1817 return EINVAL; 1818 1819 switch (cmd) { 1820 case TIOCSPGRP: 1821 if (id < 0) 1822 return EINVAL; 1823 id = -id; 1824 break; 1825 default: 1826 break; 1827 } 1828 if (id > 0) { 1829 mutex_enter(&proc_lock); 1830 error = proc_find(id) ? 0 : ESRCH; 1831 mutex_exit(&proc_lock); 1832 } else if (id < 0) { 1833 error = pgid_in_session(curproc, -id); 1834 } else { 1835 error = 0; 1836 } 1837 if (!error) { 1838 *pgid = id; 1839 } 1840 return error; 1841 } 1842 1843 void 1844 fd_set_exclose(struct lwp *l, int fd, bool exclose) 1845 { 1846 filedesc_t *fdp = l->l_fd; 1847 fdfile_t *ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd]; 1848 1849 ff->ff_exclose = exclose; 1850 if (exclose) 1851 fdp->fd_exclose = true; 1852 } 1853 1854 /* 1855 * Return descriptor owner information. If the value is positive, 1856 * it's process ID. If it's negative, it's process group ID and 1857 * needs the sign removed before use. 1858 */ 1859 int 1860 fgetown(pid_t pgid, u_long cmd, void *data) 1861 { 1862 1863 switch (cmd) { 1864 case TIOCGPGRP: 1865 *(int *)data = -pgid; 1866 break; 1867 default: 1868 *(int *)data = pgid; 1869 break; 1870 } 1871 return 0; 1872 } 1873 1874 /* 1875 * Send signal to descriptor owner, either process or process group. 1876 */ 1877 void 1878 fownsignal(pid_t pgid, int signo, int code, int band, void *fdescdata) 1879 { 1880 ksiginfo_t ksi; 1881 1882 KASSERT(!cpu_intr_p()); 1883 1884 if (pgid == 0) { 1885 return; 1886 } 1887 1888 KSI_INIT(&ksi); 1889 ksi.ksi_signo = signo; 1890 ksi.ksi_code = code; 1891 ksi.ksi_band = band; 1892 1893 mutex_enter(&proc_lock); 1894 if (pgid > 0) { 1895 struct proc *p1; 1896 1897 p1 = proc_find(pgid); 1898 if (p1 != NULL) { 1899 kpsignal(p1, &ksi, fdescdata); 1900 } 1901 } else { 1902 struct pgrp *pgrp; 1903 1904 KASSERT(pgid < 0); 1905 pgrp = pgrp_find(-pgid); 1906 if (pgrp != NULL) { 1907 kpgsignal(pgrp, &ksi, fdescdata, 0); 1908 } 1909 } 1910 mutex_exit(&proc_lock); 1911 } 1912 1913 int 1914 fd_clone(file_t *fp, unsigned fd, int flag, const struct fileops *fops, 1915 void *data) 1916 { 1917 fdfile_t *ff; 1918 filedesc_t *fdp; 1919 1920 fp->f_flag = flag & FMASK; 1921 fdp = curproc->p_fd; 1922 ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd]; 1923 KASSERT(ff != NULL); 1924 ff->ff_exclose = (flag & O_CLOEXEC) != 0; 1925 fp->f_type = DTYPE_MISC; 1926 fp->f_ops = fops; 1927 fp->f_data = data; 1928 curlwp->l_dupfd = fd; 1929 fd_affix(curproc, fp, fd); 1930 1931 return EMOVEFD; 1932 } 1933 1934 int 1935 fnullop_fcntl(file_t *fp, u_int cmd, void *data) 1936 { 1937 1938 if (cmd == F_SETFL) 1939 return 0; 1940 1941 return EOPNOTSUPP; 1942 } 1943 1944 int 1945 fnullop_poll(file_t *fp, int which) 1946 { 1947 1948 return 0; 1949 } 1950 1951 int 1952 fnullop_kqfilter(file_t *fp, struct knote *kn) 1953 { 1954 1955 return EOPNOTSUPP; 1956 } 1957 1958 void 1959 fnullop_restart(file_t *fp) 1960 { 1961 1962 } 1963 1964 int 1965 fbadop_read(file_t *fp, off_t *offset, struct uio *uio, 1966 kauth_cred_t cred, int flags) 1967 { 1968 1969 return EOPNOTSUPP; 1970 } 1971 1972 int 1973 fbadop_write(file_t *fp, off_t *offset, struct uio *uio, 1974 kauth_cred_t cred, int flags) 1975 { 1976 1977 return EOPNOTSUPP; 1978 } 1979 1980 int 1981 fbadop_ioctl(file_t *fp, u_long com, void *data) 1982 { 1983 1984 return EOPNOTSUPP; 1985 } 1986 1987 int 1988 fbadop_stat(file_t *fp, struct stat *sb) 1989 { 1990 1991 return EOPNOTSUPP; 1992 } 1993 1994 int 1995 fbadop_close(file_t *fp) 1996 { 1997 1998 return EOPNOTSUPP; 1999 } 2000 2001 /* 2002 * sysctl routines pertaining to file descriptors 2003 */ 2004 2005 /* Initialized in sysctl_init() for now... */ 2006 extern kmutex_t sysctl_file_marker_lock; 2007 static u_int sysctl_file_marker = 1; 2008 2009 /* 2010 * Expects to be called with proc_lock and sysctl_file_marker_lock locked. 2011 */ 2012 static void 2013 sysctl_file_marker_reset(void) 2014 { 2015 struct proc *p; 2016 2017 PROCLIST_FOREACH(p, &allproc) { 2018 struct filedesc *fd = p->p_fd; 2019 fdtab_t *dt; 2020 u_int i; 2021 2022 mutex_enter(&fd->fd_lock); 2023 dt = fd->fd_dt; 2024 for (i = 0; i < dt->dt_nfiles; i++) { 2025 struct file *fp; 2026 fdfile_t *ff; 2027 2028 if ((ff = dt->dt_ff[i]) == NULL) { 2029 continue; 2030 } 2031 if ((fp = atomic_load_consume(&ff->ff_file)) == NULL) { 2032 continue; 2033 } 2034 fp->f_marker = 0; 2035 } 2036 mutex_exit(&fd->fd_lock); 2037 } 2038 } 2039 2040 /* 2041 * sysctl helper routine for kern.file pseudo-subtree. 2042 */ 2043 static int 2044 sysctl_kern_file(SYSCTLFN_ARGS) 2045 { 2046 const bool allowaddr = get_expose_address(curproc); 2047 struct filelist flist; 2048 int error; 2049 size_t buflen; 2050 struct file *fp, fbuf; 2051 char *start, *where; 2052 struct proc *p; 2053 2054 start = where = oldp; 2055 buflen = *oldlenp; 2056 2057 if (where == NULL) { 2058 /* 2059 * overestimate by 10 files 2060 */ 2061 *oldlenp = sizeof(filehead) + (nfiles + 10) * 2062 sizeof(struct file); 2063 return 0; 2064 } 2065 2066 /* 2067 * first sysctl_copyout filehead 2068 */ 2069 if (buflen < sizeof(filehead)) { 2070 *oldlenp = 0; 2071 return 0; 2072 } 2073 sysctl_unlock(); 2074 if (allowaddr) { 2075 memcpy(&flist, &filehead, sizeof(flist)); 2076 } else { 2077 memset(&flist, 0, sizeof(flist)); 2078 } 2079 error = sysctl_copyout(l, &flist, where, sizeof(flist)); 2080 if (error) { 2081 sysctl_relock(); 2082 return error; 2083 } 2084 buflen -= sizeof(flist); 2085 where += sizeof(flist); 2086 2087 /* 2088 * followed by an array of file structures 2089 */ 2090 mutex_enter(&sysctl_file_marker_lock); 2091 mutex_enter(&proc_lock); 2092 PROCLIST_FOREACH(p, &allproc) { 2093 struct filedesc *fd; 2094 fdtab_t *dt; 2095 u_int i; 2096 2097 if (p->p_stat == SIDL) { 2098 /* skip embryonic processes */ 2099 continue; 2100 } 2101 mutex_enter(p->p_lock); 2102 error = kauth_authorize_process(l->l_cred, 2103 KAUTH_PROCESS_CANSEE, p, 2104 KAUTH_ARG(KAUTH_REQ_PROCESS_CANSEE_OPENFILES), 2105 NULL, NULL); 2106 mutex_exit(p->p_lock); 2107 if (error != 0) { 2108 /* 2109 * Don't leak kauth retval if we're silently 2110 * skipping this entry. 2111 */ 2112 error = 0; 2113 continue; 2114 } 2115 2116 /* 2117 * Grab a hold on the process. 2118 */ 2119 if (!rw_tryenter(&p->p_reflock, RW_READER)) { 2120 continue; 2121 } 2122 mutex_exit(&proc_lock); 2123 2124 fd = p->p_fd; 2125 mutex_enter(&fd->fd_lock); 2126 dt = fd->fd_dt; 2127 for (i = 0; i < dt->dt_nfiles; i++) { 2128 fdfile_t *ff; 2129 2130 if ((ff = dt->dt_ff[i]) == NULL) { 2131 continue; 2132 } 2133 if ((fp = atomic_load_consume(&ff->ff_file)) == NULL) { 2134 continue; 2135 } 2136 2137 mutex_enter(&fp->f_lock); 2138 2139 if ((fp->f_count == 0) || 2140 (fp->f_marker == sysctl_file_marker)) { 2141 mutex_exit(&fp->f_lock); 2142 continue; 2143 } 2144 2145 /* Check that we have enough space. */ 2146 if (buflen < sizeof(struct file)) { 2147 *oldlenp = where - start; 2148 mutex_exit(&fp->f_lock); 2149 error = ENOMEM; 2150 break; 2151 } 2152 2153 fill_file(&fbuf, fp); 2154 mutex_exit(&fp->f_lock); 2155 error = sysctl_copyout(l, &fbuf, where, sizeof(fbuf)); 2156 if (error) { 2157 break; 2158 } 2159 buflen -= sizeof(struct file); 2160 where += sizeof(struct file); 2161 2162 fp->f_marker = sysctl_file_marker; 2163 } 2164 mutex_exit(&fd->fd_lock); 2165 2166 /* 2167 * Release reference to process. 2168 */ 2169 mutex_enter(&proc_lock); 2170 rw_exit(&p->p_reflock); 2171 2172 if (error) 2173 break; 2174 } 2175 2176 sysctl_file_marker++; 2177 /* Reset all markers if wrapped. */ 2178 if (sysctl_file_marker == 0) { 2179 sysctl_file_marker_reset(); 2180 sysctl_file_marker++; 2181 } 2182 2183 mutex_exit(&proc_lock); 2184 mutex_exit(&sysctl_file_marker_lock); 2185 2186 *oldlenp = where - start; 2187 sysctl_relock(); 2188 return error; 2189 } 2190 2191 /* 2192 * sysctl helper function for kern.file2 2193 */ 2194 static int 2195 sysctl_kern_file2(SYSCTLFN_ARGS) 2196 { 2197 struct proc *p; 2198 struct file *fp; 2199 struct filedesc *fd; 2200 struct kinfo_file kf; 2201 char *dp; 2202 u_int i, op; 2203 size_t len, needed, elem_size, out_size; 2204 int error, arg, elem_count; 2205 fdfile_t *ff; 2206 fdtab_t *dt; 2207 2208 if (namelen == 1 && name[0] == CTL_QUERY) 2209 return sysctl_query(SYSCTLFN_CALL(rnode)); 2210 2211 if (namelen != 4) 2212 return EINVAL; 2213 2214 error = 0; 2215 dp = oldp; 2216 len = (oldp != NULL) ? *oldlenp : 0; 2217 op = name[0]; 2218 arg = name[1]; 2219 elem_size = name[2]; 2220 elem_count = name[3]; 2221 out_size = MIN(sizeof(kf), elem_size); 2222 needed = 0; 2223 2224 if (elem_size < 1 || elem_count < 0) 2225 return EINVAL; 2226 2227 switch (op) { 2228 case KERN_FILE_BYFILE: 2229 case KERN_FILE_BYPID: 2230 /* 2231 * We're traversing the process list in both cases; the BYFILE 2232 * case does additional work of keeping track of files already 2233 * looked at. 2234 */ 2235 2236 /* doesn't use arg so it must be zero */ 2237 if ((op == KERN_FILE_BYFILE) && (arg != 0)) 2238 return EINVAL; 2239 2240 if ((op == KERN_FILE_BYPID) && (arg < -1)) 2241 /* -1 means all processes */ 2242 return EINVAL; 2243 2244 sysctl_unlock(); 2245 if (op == KERN_FILE_BYFILE) 2246 mutex_enter(&sysctl_file_marker_lock); 2247 mutex_enter(&proc_lock); 2248 PROCLIST_FOREACH(p, &allproc) { 2249 if (p->p_stat == SIDL) { 2250 /* skip embryonic processes */ 2251 continue; 2252 } 2253 if (arg > 0 && p->p_pid != arg) { 2254 /* pick only the one we want */ 2255 /* XXX want 0 to mean "kernel files" */ 2256 continue; 2257 } 2258 mutex_enter(p->p_lock); 2259 error = kauth_authorize_process(l->l_cred, 2260 KAUTH_PROCESS_CANSEE, p, 2261 KAUTH_ARG(KAUTH_REQ_PROCESS_CANSEE_OPENFILES), 2262 NULL, NULL); 2263 mutex_exit(p->p_lock); 2264 if (error != 0) { 2265 /* 2266 * Don't leak kauth retval if we're silently 2267 * skipping this entry. 2268 */ 2269 error = 0; 2270 continue; 2271 } 2272 2273 /* 2274 * Grab a hold on the process. 2275 */ 2276 if (!rw_tryenter(&p->p_reflock, RW_READER)) { 2277 continue; 2278 } 2279 mutex_exit(&proc_lock); 2280 2281 fd = p->p_fd; 2282 mutex_enter(&fd->fd_lock); 2283 dt = fd->fd_dt; 2284 for (i = 0; i < dt->dt_nfiles; i++) { 2285 if ((ff = dt->dt_ff[i]) == NULL) { 2286 continue; 2287 } 2288 if ((fp = atomic_load_consume(&ff->ff_file)) == 2289 NULL) { 2290 continue; 2291 } 2292 2293 if ((op == KERN_FILE_BYFILE) && 2294 (fp->f_marker == sysctl_file_marker)) { 2295 continue; 2296 } 2297 if (len >= elem_size && elem_count > 0) { 2298 mutex_enter(&fp->f_lock); 2299 fill_file2(&kf, fp, ff, i, p->p_pid); 2300 mutex_exit(&fp->f_lock); 2301 mutex_exit(&fd->fd_lock); 2302 error = sysctl_copyout(l, 2303 &kf, dp, out_size); 2304 mutex_enter(&fd->fd_lock); 2305 if (error) 2306 break; 2307 dp += elem_size; 2308 len -= elem_size; 2309 } 2310 if (op == KERN_FILE_BYFILE) 2311 fp->f_marker = sysctl_file_marker; 2312 needed += elem_size; 2313 if (elem_count > 0 && elem_count != INT_MAX) 2314 elem_count--; 2315 } 2316 mutex_exit(&fd->fd_lock); 2317 2318 /* 2319 * Release reference to process. 2320 */ 2321 mutex_enter(&proc_lock); 2322 rw_exit(&p->p_reflock); 2323 } 2324 if (op == KERN_FILE_BYFILE) { 2325 sysctl_file_marker++; 2326 2327 /* Reset all markers if wrapped. */ 2328 if (sysctl_file_marker == 0) { 2329 sysctl_file_marker_reset(); 2330 sysctl_file_marker++; 2331 } 2332 } 2333 mutex_exit(&proc_lock); 2334 if (op == KERN_FILE_BYFILE) 2335 mutex_exit(&sysctl_file_marker_lock); 2336 sysctl_relock(); 2337 break; 2338 default: 2339 return EINVAL; 2340 } 2341 2342 if (oldp == NULL) 2343 needed += KERN_FILESLOP * elem_size; 2344 *oldlenp = needed; 2345 2346 return error; 2347 } 2348 2349 static void 2350 fill_file(struct file *fp, const struct file *fpsrc) 2351 { 2352 const bool allowaddr = get_expose_address(curproc); 2353 2354 memset(fp, 0, sizeof(*fp)); 2355 2356 fp->f_offset = fpsrc->f_offset; 2357 COND_SET_PTR(fp->f_cred, fpsrc->f_cred, allowaddr); 2358 COND_SET_CPTR(fp->f_ops, fpsrc->f_ops, allowaddr); 2359 COND_SET_STRUCT(fp->f_undata, fpsrc->f_undata, allowaddr); 2360 COND_SET_STRUCT(fp->f_list, fpsrc->f_list, allowaddr); 2361 fp->f_flag = fpsrc->f_flag; 2362 fp->f_marker = fpsrc->f_marker; 2363 fp->f_type = fpsrc->f_type; 2364 fp->f_advice = fpsrc->f_advice; 2365 fp->f_count = fpsrc->f_count; 2366 fp->f_msgcount = fpsrc->f_msgcount; 2367 fp->f_unpcount = fpsrc->f_unpcount; 2368 COND_SET_STRUCT(fp->f_unplist, fpsrc->f_unplist, allowaddr); 2369 } 2370 2371 static void 2372 fill_file2(struct kinfo_file *kp, const file_t *fp, const fdfile_t *ff, 2373 int i, pid_t pid) 2374 { 2375 const bool allowaddr = get_expose_address(curproc); 2376 2377 memset(kp, 0, sizeof(*kp)); 2378 2379 COND_SET_VALUE(kp->ki_fileaddr, PTRTOUINT64(fp), allowaddr); 2380 kp->ki_flag = fp->f_flag; 2381 kp->ki_iflags = 0; 2382 kp->ki_ftype = fp->f_type; 2383 kp->ki_count = fp->f_count; 2384 kp->ki_msgcount = fp->f_msgcount; 2385 COND_SET_VALUE(kp->ki_fucred, PTRTOUINT64(fp->f_cred), allowaddr); 2386 kp->ki_fuid = kauth_cred_geteuid(fp->f_cred); 2387 kp->ki_fgid = kauth_cred_getegid(fp->f_cred); 2388 COND_SET_VALUE(kp->ki_fops, PTRTOUINT64(fp->f_ops), allowaddr); 2389 kp->ki_foffset = fp->f_offset; 2390 COND_SET_VALUE(kp->ki_fdata, PTRTOUINT64(fp->f_data), allowaddr); 2391 2392 /* vnode information to glue this file to something */ 2393 if (fp->f_type == DTYPE_VNODE) { 2394 struct vnode *vp = fp->f_vnode; 2395 2396 COND_SET_VALUE(kp->ki_vun, PTRTOUINT64(vp->v_un.vu_socket), 2397 allowaddr); 2398 kp->ki_vsize = vp->v_size; 2399 kp->ki_vtype = vp->v_type; 2400 kp->ki_vtag = vp->v_tag; 2401 COND_SET_VALUE(kp->ki_vdata, PTRTOUINT64(vp->v_data), 2402 allowaddr); 2403 } 2404 2405 /* process information when retrieved via KERN_FILE_BYPID */ 2406 if (ff != NULL) { 2407 kp->ki_pid = pid; 2408 kp->ki_fd = i; 2409 kp->ki_ofileflags = ff->ff_exclose; 2410 kp->ki_usecount = ff->ff_refcnt; 2411 } 2412 } 2413
Indexes created Sat Oct 11 14:09:46 GMT 2025