1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2006, 2011, 2016-2017 Robert N. M. Watson 5 * All rights reserved. 6 * 7 * Portions of this software were developed by BAE Systems, the University of 8 * Cambridge Computer Laboratory, and Memorial University under DARPA/AFRL 9 * contract FA8650-15-C-7558 ("CADETS"), as part of the DARPA Transparent 10 * Computing (TC) research program. 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 24 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 31 * SUCH DAMAGE. 32 */ 33 34 /* 35 * Support for shared swap-backed anonymous memory objects via 36 * shm_open(2), shm_rename(2), and shm_unlink(2). 37 * While most of the implementation is here, vm_mmap.c contains 38 * mapping logic changes. 39 * 40 * posixshmcontrol(1) allows users to inspect the state of the memory 41 * objects. Per-uid swap resource limit controls total amount of 42 * memory that user can consume for anonymous objects, including 43 * shared. 44 */ 45 46 #include <sys/cdefs.h> 47 __FBSDID("$FreeBSD$"); 48 49 #include "opt_capsicum.h" 50 #include "opt_ktrace.h" 51 52 #include <sys/param.h> 53 #include <sys/capsicum.h> 54 #include <sys/conf.h> 55 #include <sys/fcntl.h> 56 #include <sys/file.h> 57 #include <sys/filedesc.h> 58 #include <sys/filio.h> 59 #include <sys/fnv_hash.h> 60 #include <sys/kernel.h> 61 #include <sys/limits.h> 62 #include <sys/uio.h> 63 #include <sys/signal.h> 64 #include <sys/jail.h> 65 #include <sys/ktrace.h> 66 #include <sys/lock.h> 67 #include <sys/malloc.h> 68 #include <sys/mman.h> 69 #include <sys/mutex.h> 70 #include <sys/priv.h> 71 #include <sys/proc.h> 72 #include <sys/refcount.h> 73 #include <sys/resourcevar.h> 74 #include <sys/rwlock.h> 75 #include <sys/sbuf.h> 76 #include <sys/stat.h> 77 #include <sys/syscallsubr.h> 78 #include <sys/sysctl.h> 79 #include <sys/sysproto.h> 80 #include <sys/systm.h> 81 #include <sys/sx.h> 82 #include <sys/time.h> 83 #include <sys/vnode.h> 84 #include <sys/unistd.h> 85 #include <sys/user.h> 86 87 #include <security/audit/audit.h> 88 #include <security/mac/mac_framework.h> 89 90 #include <vm/vm.h> 91 #include <vm/vm_param.h> 92 #include <vm/pmap.h> 93 #include <vm/vm_extern.h> 94 #include <vm/vm_map.h> 95 #include <vm/vm_kern.h> 96 #include <vm/vm_object.h> 97 #include <vm/vm_page.h> 98 #include <vm/vm_pageout.h> 99 #include <vm/vm_pager.h> 100 #include <vm/swap_pager.h> 101 102 struct shm_mapping { 103 char *sm_path; 104 Fnv32_t sm_fnv; 105 struct shmfd *sm_shmfd; 106 LIST_ENTRY(shm_mapping) sm_link; 107 }; 108 109 static MALLOC_DEFINE(M_SHMFD, "shmfd", "shared memory file descriptor"); 110 static LIST_HEAD(, shm_mapping) *shm_dictionary; 111 static struct sx shm_dict_lock; 112 static struct mtx shm_timestamp_lock; 113 static u_long shm_hash; 114 static struct unrhdr64 shm_ino_unr; 115 static dev_t shm_dev_ino; 116 117 #define SHM_HASH(fnv) (&shm_dictionary[(fnv) & shm_hash]) 118 119 static void shm_init(void *arg); 120 static void shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd); 121 static struct shmfd *shm_lookup(char *path, Fnv32_t fnv); 122 static int shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred); 123 static int shm_dotruncate_locked(struct shmfd *shmfd, off_t length, 124 void *rl_cookie); 125 static int shm_copyin_path(struct thread *td, const char *userpath_in, 126 char **path_out); 127 128 static fo_rdwr_t shm_read; 129 static fo_rdwr_t shm_write; 130 static fo_truncate_t shm_truncate; 131 static fo_ioctl_t shm_ioctl; 132 static fo_stat_t shm_stat; 133 static fo_close_t shm_close; 134 static fo_chmod_t shm_chmod; 135 static fo_chown_t shm_chown; 136 static fo_seek_t shm_seek; 137 static fo_fill_kinfo_t shm_fill_kinfo; 138 static fo_mmap_t shm_mmap; 139 static fo_get_seals_t shm_get_seals; 140 static fo_add_seals_t shm_add_seals; 141 static fo_fallocate_t shm_fallocate; 142 143 /* File descriptor operations. */ 144 struct fileops shm_ops = { 145 .fo_read = shm_read, 146 .fo_write = shm_write, 147 .fo_truncate = shm_truncate, 148 .fo_ioctl = shm_ioctl, 149 .fo_poll = invfo_poll, 150 .fo_kqfilter = invfo_kqfilter, 151 .fo_stat = shm_stat, 152 .fo_close = shm_close, 153 .fo_chmod = shm_chmod, 154 .fo_chown = shm_chown, 155 .fo_sendfile = vn_sendfile, 156 .fo_seek = shm_seek, 157 .fo_fill_kinfo = shm_fill_kinfo, 158 .fo_mmap = shm_mmap, 159 .fo_get_seals = shm_get_seals, 160 .fo_add_seals = shm_add_seals, 161 .fo_fallocate = shm_fallocate, 162 .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE 163 }; 164 165 FEATURE(posix_shm, "POSIX shared memory"); 166 167 static int 168 uiomove_object_page(vm_object_t obj, size_t len, struct uio *uio) 169 { 170 vm_page_t m; 171 vm_pindex_t idx; 172 size_t tlen; 173 int error, offset, rv; 174 175 idx = OFF_TO_IDX(uio->uio_offset); 176 offset = uio->uio_offset & PAGE_MASK; 177 tlen = MIN(PAGE_SIZE - offset, len); 178 179 VM_OBJECT_WLOCK(obj); 180 181 /* 182 * Read I/O without either a corresponding resident page or swap 183 * page: use zero_region. This is intended to avoid instantiating 184 * pages on read from a sparse region. 185 */ 186 if (uio->uio_rw == UIO_READ && vm_page_lookup(obj, idx) == NULL && 187 !vm_pager_has_page(obj, idx, NULL, NULL)) { 188 VM_OBJECT_WUNLOCK(obj); 189 return (uiomove(__DECONST(void *, zero_region), tlen, uio)); 190 } 191 192 /* 193 * Parallel reads of the page content from disk are prevented 194 * by exclusive busy. 195 * 196 * Although the tmpfs vnode lock is held here, it is 197 * nonetheless safe to sleep waiting for a free page. The 198 * pageout daemon does not need to acquire the tmpfs vnode 199 * lock to page out tobj's pages because tobj is a OBJT_SWAP 200 * type object. 201 */ 202 rv = vm_page_grab_valid(&m, obj, idx, 203 VM_ALLOC_NORMAL | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY); 204 if (rv != VM_PAGER_OK) { 205 VM_OBJECT_WUNLOCK(obj); 206 printf("uiomove_object: vm_obj %p idx %jd pager error %d\n", 207 obj, idx, rv); 208 return (EIO); 209 } 210 VM_OBJECT_WUNLOCK(obj); 211 error = uiomove_fromphys(&m, offset, tlen, uio); 212 if (uio->uio_rw == UIO_WRITE && error == 0) 213 vm_page_set_dirty(m); 214 vm_page_activate(m); 215 vm_page_sunbusy(m); 216 217 return (error); 218 } 219 220 int 221 uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio) 222 { 223 ssize_t resid; 224 size_t len; 225 int error; 226 227 error = 0; 228 while ((resid = uio->uio_resid) > 0) { 229 if (obj_size <= uio->uio_offset) 230 break; 231 len = MIN(obj_size - uio->uio_offset, resid); 232 if (len == 0) 233 break; 234 error = uiomove_object_page(obj, len, uio); 235 if (error != 0 || resid == uio->uio_resid) 236 break; 237 } 238 return (error); 239 } 240 241 static int 242 shm_seek(struct file *fp, off_t offset, int whence, struct thread *td) 243 { 244 struct shmfd *shmfd; 245 off_t foffset; 246 int error; 247 248 shmfd = fp->f_data; 249 foffset = foffset_lock(fp, 0); 250 error = 0; 251 switch (whence) { 252 case L_INCR: 253 if (foffset < 0 || 254 (offset > 0 && foffset > OFF_MAX - offset)) { 255 error = EOVERFLOW; 256 break; 257 } 258 offset += foffset; 259 break; 260 case L_XTND: 261 if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) { 262 error = EOVERFLOW; 263 break; 264 } 265 offset += shmfd->shm_size; 266 break; 267 case L_SET: 268 break; 269 default: 270 error = EINVAL; 271 } 272 if (error == 0) { 273 if (offset < 0 || offset > shmfd->shm_size) 274 error = EINVAL; 275 else 276 td->td_uretoff.tdu_off = offset; 277 } 278 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0); 279 return (error); 280 } 281 282 static int 283 shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred, 284 int flags, struct thread *td) 285 { 286 struct shmfd *shmfd; 287 void *rl_cookie; 288 int error; 289 290 shmfd = fp->f_data; 291 #ifdef MAC 292 error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd); 293 if (error) 294 return (error); 295 #endif 296 foffset_lock_uio(fp, uio, flags); 297 rl_cookie = rangelock_rlock(&shmfd->shm_rl, uio->uio_offset, 298 uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx); 299 error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio); 300 rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); 301 foffset_unlock_uio(fp, uio, flags); 302 return (error); 303 } 304 305 static int 306 shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred, 307 int flags, struct thread *td) 308 { 309 struct shmfd *shmfd; 310 void *rl_cookie; 311 int error; 312 313 shmfd = fp->f_data; 314 #ifdef MAC 315 error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd); 316 if (error) 317 return (error); 318 #endif 319 foffset_lock_uio(fp, uio, flags); 320 if ((flags & FOF_OFFSET) == 0) { 321 rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX, 322 &shmfd->shm_mtx); 323 } else { 324 rl_cookie = rangelock_wlock(&shmfd->shm_rl, uio->uio_offset, 325 uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx); 326 } 327 if ((shmfd->shm_seals & F_SEAL_WRITE) != 0) 328 error = EPERM; 329 else 330 error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio); 331 rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); 332 foffset_unlock_uio(fp, uio, flags); 333 return (error); 334 } 335 336 static int 337 shm_truncate(struct file *fp, off_t length, struct ucred *active_cred, 338 struct thread *td) 339 { 340 struct shmfd *shmfd; 341 #ifdef MAC 342 int error; 343 #endif 344 345 shmfd = fp->f_data; 346 #ifdef MAC 347 error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd); 348 if (error) 349 return (error); 350 #endif 351 return (shm_dotruncate(shmfd, length)); 352 } 353 354 int 355 shm_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred, 356 struct thread *td) 357 { 358 359 switch (com) { 360 case FIONBIO: 361 case FIOASYNC: 362 /* 363 * Allow fcntl(fd, F_SETFL, O_NONBLOCK) to work, 364 * just like it would on an unlinked regular file 365 */ 366 return (0); 367 default: 368 return (ENOTTY); 369 } 370 } 371 372 static int 373 shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred, 374 struct thread *td) 375 { 376 struct shmfd *shmfd; 377 #ifdef MAC 378 int error; 379 #endif 380 381 shmfd = fp->f_data; 382 383 #ifdef MAC 384 error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd); 385 if (error) 386 return (error); 387 #endif 388 389 /* 390 * Attempt to return sanish values for fstat() on a memory file 391 * descriptor. 392 */ 393 bzero(sb, sizeof(*sb)); 394 sb->st_blksize = PAGE_SIZE; 395 sb->st_size = shmfd->shm_size; 396 sb->st_blocks = howmany(sb->st_size, sb->st_blksize); 397 mtx_lock(&shm_timestamp_lock); 398 sb->st_atim = shmfd->shm_atime; 399 sb->st_ctim = shmfd->shm_ctime; 400 sb->st_mtim = shmfd->shm_mtime; 401 sb->st_birthtim = shmfd->shm_birthtime; 402 sb->st_mode = S_IFREG | shmfd->shm_mode; /* XXX */ 403 sb->st_uid = shmfd->shm_uid; 404 sb->st_gid = shmfd->shm_gid; 405 mtx_unlock(&shm_timestamp_lock); 406 sb->st_dev = shm_dev_ino; 407 sb->st_ino = shmfd->shm_ino; 408 sb->st_nlink = shmfd->shm_object->ref_count; 409 410 return (0); 411 } 412 413 static int 414 shm_close(struct file *fp, struct thread *td) 415 { 416 struct shmfd *shmfd; 417 418 shmfd = fp->f_data; 419 fp->f_data = NULL; 420 shm_drop(shmfd); 421 422 return (0); 423 } 424 425 static int 426 shm_copyin_path(struct thread *td, const char *userpath_in, char **path_out) { 427 int error; 428 char *path; 429 const char *pr_path; 430 size_t pr_pathlen; 431 432 path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK); 433 pr_path = td->td_ucred->cr_prison->pr_path; 434 435 /* Construct a full pathname for jailed callers. */ 436 pr_pathlen = strcmp(pr_path, "/") == 437 0 ? 0 : strlcpy(path, pr_path, MAXPATHLEN); 438 error = copyinstr(userpath_in, path + pr_pathlen, 439 MAXPATHLEN - pr_pathlen, NULL); 440 if (error != 0) 441 goto out; 442 443 #ifdef KTRACE 444 if (KTRPOINT(curthread, KTR_NAMEI)) 445 ktrnamei(path); 446 #endif 447 448 /* Require paths to start with a '/' character. */ 449 if (path[pr_pathlen] != '/') { 450 error = EINVAL; 451 goto out; 452 } 453 454 *path_out = path; 455 456 out: 457 if (error != 0) 458 free(path, M_SHMFD); 459 460 return (error); 461 } 462 463 static int 464 shm_dotruncate_locked(struct shmfd *shmfd, off_t length, void *rl_cookie) 465 { 466 vm_object_t object; 467 vm_page_t m; 468 vm_pindex_t idx, nobjsize; 469 vm_ooffset_t delta; 470 int base, rv; 471 472 KASSERT(length >= 0, ("shm_dotruncate: length < 0")); 473 object = shmfd->shm_object; 474 VM_OBJECT_ASSERT_WLOCKED(object); 475 rangelock_cookie_assert(rl_cookie, RA_WLOCKED); 476 if (length == shmfd->shm_size) 477 return (0); 478 nobjsize = OFF_TO_IDX(length + PAGE_MASK); 479 480 /* Are we shrinking? If so, trim the end. */ 481 if (length < shmfd->shm_size) { 482 if ((shmfd->shm_seals & F_SEAL_SHRINK) != 0) 483 return (EPERM); 484 485 /* 486 * Disallow any requests to shrink the size if this 487 * object is mapped into the kernel. 488 */ 489 if (shmfd->shm_kmappings > 0) 490 return (EBUSY); 491 492 /* 493 * Zero the truncated part of the last page. 494 */ 495 base = length & PAGE_MASK; 496 if (base != 0) { 497 idx = OFF_TO_IDX(length); 498 retry: 499 m = vm_page_grab(object, idx, VM_ALLOC_NOCREAT); 500 if (m != NULL) { 501 MPASS(vm_page_all_valid(m)); 502 } else if (vm_pager_has_page(object, idx, NULL, NULL)) { 503 m = vm_page_alloc(object, idx, 504 VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL); 505 if (m == NULL) 506 goto retry; 507 vm_object_pip_add(object, 1); 508 VM_OBJECT_WUNLOCK(object); 509 rv = vm_pager_get_pages(object, &m, 1, NULL, 510 NULL); 511 VM_OBJECT_WLOCK(object); 512 vm_object_pip_wakeup(object); 513 if (rv == VM_PAGER_OK) { 514 /* 515 * Since the page was not resident, 516 * and therefore not recently 517 * accessed, immediately enqueue it 518 * for asynchronous laundering. The 519 * current operation is not regarded 520 * as an access. 521 */ 522 vm_page_launder(m); 523 } else { 524 vm_page_free(m); 525 VM_OBJECT_WUNLOCK(object); 526 return (EIO); 527 } 528 } 529 if (m != NULL) { 530 pmap_zero_page_area(m, base, PAGE_SIZE - base); 531 KASSERT(vm_page_all_valid(m), 532 ("shm_dotruncate: page %p is invalid", m)); 533 vm_page_set_dirty(m); 534 vm_page_xunbusy(m); 535 } 536 } 537 delta = IDX_TO_OFF(object->size - nobjsize); 538 539 /* Toss in memory pages. */ 540 if (nobjsize < object->size) 541 vm_object_page_remove(object, nobjsize, object->size, 542 0); 543 544 /* Toss pages from swap. */ 545 if (object->type == OBJT_SWAP) 546 swap_pager_freespace(object, nobjsize, delta); 547 548 /* Free the swap accounted for shm */ 549 swap_release_by_cred(delta, object->cred); 550 object->charge -= delta; 551 } else { 552 if ((shmfd->shm_seals & F_SEAL_GROW) != 0) 553 return (EPERM); 554 555 /* Try to reserve additional swap space. */ 556 delta = IDX_TO_OFF(nobjsize - object->size); 557 if (!swap_reserve_by_cred(delta, object->cred)) 558 return (ENOMEM); 559 object->charge += delta; 560 } 561 shmfd->shm_size = length; 562 mtx_lock(&shm_timestamp_lock); 563 vfs_timestamp(&shmfd->shm_ctime); 564 shmfd->shm_mtime = shmfd->shm_ctime; 565 mtx_unlock(&shm_timestamp_lock); 566 object->size = nobjsize; 567 return (0); 568 } 569 570 int 571 shm_dotruncate(struct shmfd *shmfd, off_t length) 572 { 573 void *rl_cookie; 574 int error; 575 576 rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX, 577 &shmfd->shm_mtx); 578 VM_OBJECT_WLOCK(shmfd->shm_object); 579 error = shm_dotruncate_locked(shmfd, length, rl_cookie); 580 VM_OBJECT_WUNLOCK(shmfd->shm_object); 581 rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); 582 return (error); 583 } 584 585 /* 586 * shmfd object management including creation and reference counting 587 * routines. 588 */ 589 struct shmfd * 590 shm_alloc(struct ucred *ucred, mode_t mode) 591 { 592 struct shmfd *shmfd; 593 594 shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO); 595 shmfd->shm_size = 0; 596 shmfd->shm_uid = ucred->cr_uid; 597 shmfd->shm_gid = ucred->cr_gid; 598 shmfd->shm_mode = mode; 599 shmfd->shm_object = vm_pager_allocate(OBJT_SWAP, NULL, 600 shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred); 601 KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate")); 602 vfs_timestamp(&shmfd->shm_birthtime); 603 shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime = 604 shmfd->shm_birthtime; 605 shmfd->shm_ino = alloc_unr64(&shm_ino_unr); 606 refcount_init(&shmfd->shm_refs, 1); 607 mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF); 608 rangelock_init(&shmfd->shm_rl); 609 #ifdef MAC 610 mac_posixshm_init(shmfd); 611 mac_posixshm_create(ucred, shmfd); 612 #endif 613 614 return (shmfd); 615 } 616 617 struct shmfd * 618 shm_hold(struct shmfd *shmfd) 619 { 620 621 refcount_acquire(&shmfd->shm_refs); 622 return (shmfd); 623 } 624 625 void 626 shm_drop(struct shmfd *shmfd) 627 { 628 629 if (refcount_release(&shmfd->shm_refs)) { 630 #ifdef MAC 631 mac_posixshm_destroy(shmfd); 632 #endif 633 rangelock_destroy(&shmfd->shm_rl); 634 mtx_destroy(&shmfd->shm_mtx); 635 vm_object_deallocate(shmfd->shm_object); 636 free(shmfd, M_SHMFD); 637 } 638 } 639 640 /* 641 * Determine if the credentials have sufficient permissions for a 642 * specified combination of FREAD and FWRITE. 643 */ 644 int 645 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags) 646 { 647 accmode_t accmode; 648 int error; 649 650 accmode = 0; 651 if (flags & FREAD) 652 accmode |= VREAD; 653 if (flags & FWRITE) 654 accmode |= VWRITE; 655 mtx_lock(&shm_timestamp_lock); 656 error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid, 657 accmode, ucred, NULL); 658 mtx_unlock(&shm_timestamp_lock); 659 return (error); 660 } 661 662 /* 663 * Dictionary management. We maintain an in-kernel dictionary to map 664 * paths to shmfd objects. We use the FNV hash on the path to store 665 * the mappings in a hash table. 666 */ 667 static void 668 shm_init(void *arg) 669 { 670 671 mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF); 672 sx_init(&shm_dict_lock, "shm dictionary"); 673 shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash); 674 new_unrhdr64(&shm_ino_unr, 1); 675 shm_dev_ino = devfs_alloc_cdp_inode(); 676 KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized")); 677 } 678 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL); 679 680 static struct shmfd * 681 shm_lookup(char *path, Fnv32_t fnv) 682 { 683 struct shm_mapping *map; 684 685 LIST_FOREACH(map, SHM_HASH(fnv), sm_link) { 686 if (map->sm_fnv != fnv) 687 continue; 688 if (strcmp(map->sm_path, path) == 0) 689 return (map->sm_shmfd); 690 } 691 692 return (NULL); 693 } 694 695 static void 696 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd) 697 { 698 struct shm_mapping *map; 699 700 map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK); 701 map->sm_path = path; 702 map->sm_fnv = fnv; 703 map->sm_shmfd = shm_hold(shmfd); 704 shmfd->shm_path = path; 705 LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link); 706 } 707 708 static int 709 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred) 710 { 711 struct shm_mapping *map; 712 int error; 713 714 LIST_FOREACH(map, SHM_HASH(fnv), sm_link) { 715 if (map->sm_fnv != fnv) 716 continue; 717 if (strcmp(map->sm_path, path) == 0) { 718 #ifdef MAC 719 error = mac_posixshm_check_unlink(ucred, map->sm_shmfd); 720 if (error) 721 return (error); 722 #endif 723 error = shm_access(map->sm_shmfd, ucred, 724 FREAD | FWRITE); 725 if (error) 726 return (error); 727 map->sm_shmfd->shm_path = NULL; 728 LIST_REMOVE(map, sm_link); 729 shm_drop(map->sm_shmfd); 730 free(map->sm_path, M_SHMFD); 731 free(map, M_SHMFD); 732 return (0); 733 } 734 } 735 736 return (ENOENT); 737 } 738 739 int 740 kern_shm_open2(struct thread *td, const char *userpath, int flags, mode_t mode, 741 int shmflags, struct filecaps *fcaps, const char *name __unused) 742 { 743 struct filedesc *fdp; 744 struct shmfd *shmfd; 745 struct file *fp; 746 char *path; 747 void *rl_cookie; 748 Fnv32_t fnv; 749 mode_t cmode; 750 int error, fd, initial_seals; 751 752 if ((shmflags & ~SHM_ALLOW_SEALING) != 0) 753 return (EINVAL); 754 755 initial_seals = F_SEAL_SEAL; 756 if ((shmflags & SHM_ALLOW_SEALING) != 0) 757 initial_seals &= ~F_SEAL_SEAL; 758 759 #ifdef CAPABILITY_MODE 760 /* 761 * shm_open(2) is only allowed for anonymous objects. 762 */ 763 if (IN_CAPABILITY_MODE(td) && (userpath != SHM_ANON)) 764 return (ECAPMODE); 765 #endif 766 767 AUDIT_ARG_FFLAGS(flags); 768 AUDIT_ARG_MODE(mode); 769 770 if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR) 771 return (EINVAL); 772 773 if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0) 774 return (EINVAL); 775 776 /* 777 * Currently only F_SEAL_SEAL may be set when creating or opening shmfd. 778 * If the decision is made later to allow additional seals, care must be 779 * taken below to ensure that the seals are properly set if the shmfd 780 * already existed -- this currently assumes that only F_SEAL_SEAL can 781 * be set and doesn't take further precautions to ensure the validity of 782 * the seals being added with respect to current mappings. 783 */ 784 if ((initial_seals & ~F_SEAL_SEAL) != 0) 785 return (EINVAL); 786 787 fdp = td->td_proc->p_fd; 788 cmode = (mode & ~fdp->fd_cmask) & ACCESSPERMS; 789 790 /* 791 * shm_open(2) created shm should always have O_CLOEXEC set, as mandated 792 * by POSIX. We allow it to be unset here so that an in-kernel 793 * interface may be written as a thin layer around shm, optionally not 794 * setting CLOEXEC. For shm_open(2), O_CLOEXEC is set unconditionally 795 * in sys_shm_open() to keep this implementation compliant. 796 */ 797 error = falloc_caps(td, &fp, &fd, flags & O_CLOEXEC, fcaps); 798 if (error) 799 return (error); 800 801 /* A SHM_ANON path pointer creates an anonymous object. */ 802 if (userpath == SHM_ANON) { 803 /* A read-only anonymous object is pointless. */ 804 if ((flags & O_ACCMODE) == O_RDONLY) { 805 fdclose(td, fp, fd); 806 fdrop(fp, td); 807 return (EINVAL); 808 } 809 shmfd = shm_alloc(td->td_ucred, cmode); 810 shmfd->shm_seals = initial_seals; 811 } else { 812 error = shm_copyin_path(td, userpath, &path); 813 if (error != 0) { 814 fdclose(td, fp, fd); 815 fdrop(fp, td); 816 return (error); 817 } 818 819 AUDIT_ARG_UPATH1_CANON(path); 820 fnv = fnv_32_str(path, FNV1_32_INIT); 821 sx_xlock(&shm_dict_lock); 822 shmfd = shm_lookup(path, fnv); 823 if (shmfd == NULL) { 824 /* Object does not yet exist, create it if requested. */ 825 if (flags & O_CREAT) { 826 #ifdef MAC 827 error = mac_posixshm_check_create(td->td_ucred, 828 path); 829 if (error == 0) { 830 #endif 831 shmfd = shm_alloc(td->td_ucred, cmode); 832 shmfd->shm_seals = initial_seals; 833 shm_insert(path, fnv, shmfd); 834 #ifdef MAC 835 } 836 #endif 837 } else { 838 free(path, M_SHMFD); 839 error = ENOENT; 840 } 841 } else { 842 rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX, 843 &shmfd->shm_mtx); 844 845 /* 846 * kern_shm_open() likely shouldn't ever error out on 847 * trying to set a seal that already exists, unlike 848 * F_ADD_SEALS. This would break terribly as 849 * shm_open(2) actually sets F_SEAL_SEAL to maintain 850 * historical behavior where the underlying file could 851 * not be sealed. 852 */ 853 initial_seals &= ~shmfd->shm_seals; 854 855 /* 856 * Object already exists, obtain a new 857 * reference if requested and permitted. 858 */ 859 free(path, M_SHMFD); 860 861 /* 862 * initial_seals can't set additional seals if we've 863 * already been set F_SEAL_SEAL. If F_SEAL_SEAL is set, 864 * then we've already removed that one from 865 * initial_seals. This is currently redundant as we 866 * only allow setting F_SEAL_SEAL at creation time, but 867 * it's cheap to check and decreases the effort required 868 * to allow additional seals. 869 */ 870 if ((shmfd->shm_seals & F_SEAL_SEAL) != 0 && 871 initial_seals != 0) 872 error = EPERM; 873 else if ((flags & (O_CREAT | O_EXCL)) == 874 (O_CREAT | O_EXCL)) 875 error = EEXIST; 876 else { 877 #ifdef MAC 878 error = mac_posixshm_check_open(td->td_ucred, 879 shmfd, FFLAGS(flags & O_ACCMODE)); 880 if (error == 0) 881 #endif 882 error = shm_access(shmfd, td->td_ucred, 883 FFLAGS(flags & O_ACCMODE)); 884 } 885 886 /* 887 * Truncate the file back to zero length if 888 * O_TRUNC was specified and the object was 889 * opened with read/write. 890 */ 891 if (error == 0 && 892 (flags & (O_ACCMODE | O_TRUNC)) == 893 (O_RDWR | O_TRUNC)) { 894 VM_OBJECT_WLOCK(shmfd->shm_object); 895 #ifdef MAC 896 error = mac_posixshm_check_truncate( 897 td->td_ucred, fp->f_cred, shmfd); 898 if (error == 0) 899 #endif 900 error = shm_dotruncate_locked(shmfd, 0, 901 rl_cookie); 902 VM_OBJECT_WUNLOCK(shmfd->shm_object); 903 } 904 if (error == 0) { 905 /* 906 * Currently we only allow F_SEAL_SEAL to be 907 * set initially. As noted above, this would 908 * need to be reworked should that change. 909 */ 910 shmfd->shm_seals |= initial_seals; 911 shm_hold(shmfd); 912 } 913 rangelock_unlock(&shmfd->shm_rl, rl_cookie, 914 &shmfd->shm_mtx); 915 } 916 sx_xunlock(&shm_dict_lock); 917 918 if (error) { 919 fdclose(td, fp, fd); 920 fdrop(fp, td); 921 return (error); 922 } 923 } 924 925 finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops); 926 927 td->td_retval[0] = fd; 928 fdrop(fp, td); 929 930 return (0); 931 } 932 933 /* System calls. */ 934 #ifdef COMPAT_FREEBSD12 935 int 936 freebsd12_shm_open(struct thread *td, struct freebsd12_shm_open_args *uap) 937 { 938 939 return (kern_shm_open(td, uap->path, uap->flags | O_CLOEXEC, 940 uap->mode, NULL)); 941 } 942 #endif 943 944 int 945 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap) 946 { 947 char *path; 948 Fnv32_t fnv; 949 int error; 950 951 error = shm_copyin_path(td, uap->path, &path); 952 if (error != 0) 953 return (error); 954 955 AUDIT_ARG_UPATH1_CANON(path); 956 fnv = fnv_32_str(path, FNV1_32_INIT); 957 sx_xlock(&shm_dict_lock); 958 error = shm_remove(path, fnv, td->td_ucred); 959 sx_xunlock(&shm_dict_lock); 960 free(path, M_TEMP); 961 962 return (error); 963 } 964 965 int 966 sys_shm_rename(struct thread *td, struct shm_rename_args *uap) 967 { 968 char *path_from = NULL, *path_to = NULL; 969 Fnv32_t fnv_from, fnv_to; 970 struct shmfd *fd_from; 971 struct shmfd *fd_to; 972 int error; 973 int flags; 974 975 flags = uap->flags; 976 AUDIT_ARG_FFLAGS(flags); 977 978 /* 979 * Make sure the user passed only valid flags. 980 * If you add a new flag, please add a new term here. 981 */ 982 if ((flags & ~( 983 SHM_RENAME_NOREPLACE | 984 SHM_RENAME_EXCHANGE 985 )) != 0) { 986 error = EINVAL; 987 goto out; 988 } 989 990 /* 991 * EXCHANGE and NOREPLACE don't quite make sense together. Let's 992 * force the user to choose one or the other. 993 */ 994 if ((flags & SHM_RENAME_NOREPLACE) != 0 && 995 (flags & SHM_RENAME_EXCHANGE) != 0) { 996 error = EINVAL; 997 goto out; 998 } 999 1000 /* Renaming to or from anonymous makes no sense */ 1001 if (uap->path_from == SHM_ANON || uap->path_to == SHM_ANON) { 1002 error = EINVAL; 1003 goto out; 1004 } 1005 1006 error = shm_copyin_path(td, uap->path_from, &path_from); 1007 if (error != 0) 1008 goto out; 1009 1010 error = shm_copyin_path(td, uap->path_to, &path_to); 1011 if (error != 0) 1012 goto out; 1013 1014 AUDIT_ARG_UPATH1_CANON(path_from); 1015 AUDIT_ARG_UPATH2_CANON(path_to); 1016 1017 /* Rename with from/to equal is a no-op */ 1018 if (strcmp(path_from, path_to) == 0) 1019 goto out; 1020 1021 fnv_from = fnv_32_str(path_from, FNV1_32_INIT); 1022 fnv_to = fnv_32_str(path_to, FNV1_32_INIT); 1023 1024 sx_xlock(&shm_dict_lock); 1025 1026 fd_from = shm_lookup(path_from, fnv_from); 1027 if (fd_from == NULL) { 1028 error = ENOENT; 1029 goto out_locked; 1030 } 1031 1032 fd_to = shm_lookup(path_to, fnv_to); 1033 if ((flags & SHM_RENAME_NOREPLACE) != 0 && fd_to != NULL) { 1034 error = EEXIST; 1035 goto out_locked; 1036 } 1037 1038 /* 1039 * Unconditionally prevents shm_remove from invalidating the 'from' 1040 * shm's state. 1041 */ 1042 shm_hold(fd_from); 1043 error = shm_remove(path_from, fnv_from, td->td_ucred); 1044 1045 /* 1046 * One of my assumptions failed if ENOENT (e.g. locking didn't 1047 * protect us) 1048 */ 1049 KASSERT(error != ENOENT, ("Our shm disappeared during shm_rename: %s", 1050 path_from)); 1051 if (error != 0) { 1052 shm_drop(fd_from); 1053 goto out_locked; 1054 } 1055 1056 /* 1057 * If we are exchanging, we need to ensure the shm_remove below 1058 * doesn't invalidate the dest shm's state. 1059 */ 1060 if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) 1061 shm_hold(fd_to); 1062 1063 /* 1064 * NOTE: if path_to is not already in the hash, c'est la vie; 1065 * it simply means we have nothing already at path_to to unlink. 1066 * That is the ENOENT case. 1067 * 1068 * If we somehow don't have access to unlink this guy, but 1069 * did for the shm at path_from, then relink the shm to path_from 1070 * and abort with EACCES. 1071 * 1072 * All other errors: that is weird; let's relink and abort the 1073 * operation. 1074 */ 1075 error = shm_remove(path_to, fnv_to, td->td_ucred); 1076 if (error != 0 && error != ENOENT) { 1077 shm_insert(path_from, fnv_from, fd_from); 1078 shm_drop(fd_from); 1079 /* Don't free path_from now, since the hash references it */ 1080 path_from = NULL; 1081 goto out_locked; 1082 } 1083 1084 error = 0; 1085 1086 shm_insert(path_to, fnv_to, fd_from); 1087 1088 /* Don't free path_to now, since the hash references it */ 1089 path_to = NULL; 1090 1091 /* We kept a ref when we removed, and incremented again in insert */ 1092 shm_drop(fd_from); 1093 KASSERT(fd_from->shm_refs > 0, ("Expected >0 refs; got: %d\n", 1094 fd_from->shm_refs)); 1095 1096 if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) { 1097 shm_insert(path_from, fnv_from, fd_to); 1098 path_from = NULL; 1099 shm_drop(fd_to); 1100 KASSERT(fd_to->shm_refs > 0, ("Expected >0 refs; got: %d\n", 1101 fd_to->shm_refs)); 1102 } 1103 1104 out_locked: 1105 sx_xunlock(&shm_dict_lock); 1106 1107 out: 1108 free(path_from, M_SHMFD); 1109 free(path_to, M_SHMFD); 1110 return (error); 1111 } 1112 1113 int 1114 shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize, 1115 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, 1116 vm_ooffset_t foff, struct thread *td) 1117 { 1118 struct shmfd *shmfd; 1119 vm_prot_t maxprot; 1120 int error; 1121 bool writecnt; 1122 void *rl_cookie; 1123 1124 shmfd = fp->f_data; 1125 maxprot = VM_PROT_NONE; 1126 1127 rl_cookie = rangelock_rlock(&shmfd->shm_rl, 0, objsize, 1128 &shmfd->shm_mtx); 1129 /* FREAD should always be set. */ 1130 if ((fp->f_flag & FREAD) != 0) 1131 maxprot |= VM_PROT_EXECUTE | VM_PROT_READ; 1132 1133 /* 1134 * If FWRITE's set, we can allow VM_PROT_WRITE unless it's a shared 1135 * mapping with a write seal applied. 1136 */ 1137 if ((fp->f_flag & FWRITE) != 0 && ((flags & MAP_SHARED) == 0 || 1138 (shmfd->shm_seals & F_SEAL_WRITE) == 0)) 1139 maxprot |= VM_PROT_WRITE; 1140 1141 writecnt = (flags & MAP_SHARED) != 0 && (prot & VM_PROT_WRITE) != 0; 1142 1143 if (writecnt && (shmfd->shm_seals & F_SEAL_WRITE) != 0) { 1144 error = EPERM; 1145 goto out; 1146 } 1147 1148 /* Don't permit shared writable mappings on read-only descriptors. */ 1149 if (writecnt && (maxprot & VM_PROT_WRITE) == 0) { 1150 error = EACCES; 1151 goto out; 1152 } 1153 maxprot &= cap_maxprot; 1154 1155 /* See comment in vn_mmap(). */ 1156 if ( 1157 #ifdef _LP64 1158 objsize > OFF_MAX || 1159 #endif 1160 foff < 0 || foff > OFF_MAX - objsize) { 1161 error = EINVAL; 1162 goto out; 1163 } 1164 1165 #ifdef MAC 1166 error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags); 1167 if (error != 0) 1168 goto out; 1169 #endif 1170 1171 mtx_lock(&shm_timestamp_lock); 1172 vfs_timestamp(&shmfd->shm_atime); 1173 mtx_unlock(&shm_timestamp_lock); 1174 vm_object_reference(shmfd->shm_object); 1175 1176 if (writecnt) 1177 vm_pager_update_writecount(shmfd->shm_object, 0, objsize); 1178 error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags, 1179 shmfd->shm_object, foff, writecnt, td); 1180 if (error != 0) { 1181 if (writecnt) 1182 vm_pager_release_writecount(shmfd->shm_object, 0, 1183 objsize); 1184 vm_object_deallocate(shmfd->shm_object); 1185 } 1186 out: 1187 rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); 1188 return (error); 1189 } 1190 1191 static int 1192 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, 1193 struct thread *td) 1194 { 1195 struct shmfd *shmfd; 1196 int error; 1197 1198 error = 0; 1199 shmfd = fp->f_data; 1200 mtx_lock(&shm_timestamp_lock); 1201 /* 1202 * SUSv4 says that x bits of permission need not be affected. 1203 * Be consistent with our shm_open there. 1204 */ 1205 #ifdef MAC 1206 error = mac_posixshm_check_setmode(active_cred, shmfd, mode); 1207 if (error != 0) 1208 goto out; 1209 #endif 1210 error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, 1211 shmfd->shm_gid, VADMIN, active_cred, NULL); 1212 if (error != 0) 1213 goto out; 1214 shmfd->shm_mode = mode & ACCESSPERMS; 1215 out: 1216 mtx_unlock(&shm_timestamp_lock); 1217 return (error); 1218 } 1219 1220 static int 1221 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred, 1222 struct thread *td) 1223 { 1224 struct shmfd *shmfd; 1225 int error; 1226 1227 error = 0; 1228 shmfd = fp->f_data; 1229 mtx_lock(&shm_timestamp_lock); 1230 #ifdef MAC 1231 error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid); 1232 if (error != 0) 1233 goto out; 1234 #endif 1235 if (uid == (uid_t)-1) 1236 uid = shmfd->shm_uid; 1237 if (gid == (gid_t)-1) 1238 gid = shmfd->shm_gid; 1239 if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) || 1240 (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) && 1241 (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN))) 1242 goto out; 1243 shmfd->shm_uid = uid; 1244 shmfd->shm_gid = gid; 1245 out: 1246 mtx_unlock(&shm_timestamp_lock); 1247 return (error); 1248 } 1249 1250 /* 1251 * Helper routines to allow the backing object of a shared memory file 1252 * descriptor to be mapped in the kernel. 1253 */ 1254 int 1255 shm_map(struct file *fp, size_t size, off_t offset, void **memp) 1256 { 1257 struct shmfd *shmfd; 1258 vm_offset_t kva, ofs; 1259 vm_object_t obj; 1260 int rv; 1261 1262 if (fp->f_type != DTYPE_SHM) 1263 return (EINVAL); 1264 shmfd = fp->f_data; 1265 obj = shmfd->shm_object; 1266 VM_OBJECT_WLOCK(obj); 1267 /* 1268 * XXXRW: This validation is probably insufficient, and subject to 1269 * sign errors. It should be fixed. 1270 */ 1271 if (offset >= shmfd->shm_size || 1272 offset + size > round_page(shmfd->shm_size)) { 1273 VM_OBJECT_WUNLOCK(obj); 1274 return (EINVAL); 1275 } 1276 1277 shmfd->shm_kmappings++; 1278 vm_object_reference_locked(obj); 1279 VM_OBJECT_WUNLOCK(obj); 1280 1281 /* Map the object into the kernel_map and wire it. */ 1282 kva = vm_map_min(kernel_map); 1283 ofs = offset & PAGE_MASK; 1284 offset = trunc_page(offset); 1285 size = round_page(size + ofs); 1286 rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0, 1287 VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE, 1288 VM_PROT_READ | VM_PROT_WRITE, 0); 1289 if (rv == KERN_SUCCESS) { 1290 rv = vm_map_wire(kernel_map, kva, kva + size, 1291 VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES); 1292 if (rv == KERN_SUCCESS) { 1293 *memp = (void *)(kva + ofs); 1294 return (0); 1295 } 1296 vm_map_remove(kernel_map, kva, kva + size); 1297 } else 1298 vm_object_deallocate(obj); 1299 1300 /* On failure, drop our mapping reference. */ 1301 VM_OBJECT_WLOCK(obj); 1302 shmfd->shm_kmappings--; 1303 VM_OBJECT_WUNLOCK(obj); 1304 1305 return (vm_mmap_to_errno(rv)); 1306 } 1307 1308 /* 1309 * We require the caller to unmap the entire entry. This allows us to 1310 * safely decrement shm_kmappings when a mapping is removed. 1311 */ 1312 int 1313 shm_unmap(struct file *fp, void *mem, size_t size) 1314 { 1315 struct shmfd *shmfd; 1316 vm_map_entry_t entry; 1317 vm_offset_t kva, ofs; 1318 vm_object_t obj; 1319 vm_pindex_t pindex; 1320 vm_prot_t prot; 1321 boolean_t wired; 1322 vm_map_t map; 1323 int rv; 1324 1325 if (fp->f_type != DTYPE_SHM) 1326 return (EINVAL); 1327 shmfd = fp->f_data; 1328 kva = (vm_offset_t)mem; 1329 ofs = kva & PAGE_MASK; 1330 kva = trunc_page(kva); 1331 size = round_page(size + ofs); 1332 map = kernel_map; 1333 rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry, 1334 &obj, &pindex, &prot, &wired); 1335 if (rv != KERN_SUCCESS) 1336 return (EINVAL); 1337 if (entry->start != kva || entry->end != kva + size) { 1338 vm_map_lookup_done(map, entry); 1339 return (EINVAL); 1340 } 1341 vm_map_lookup_done(map, entry); 1342 if (obj != shmfd->shm_object) 1343 return (EINVAL); 1344 vm_map_remove(map, kva, kva + size); 1345 VM_OBJECT_WLOCK(obj); 1346 KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped")); 1347 shmfd->shm_kmappings--; 1348 VM_OBJECT_WUNLOCK(obj); 1349 return (0); 1350 } 1351 1352 static int 1353 shm_fill_kinfo_locked(struct shmfd *shmfd, struct kinfo_file *kif, bool list) 1354 { 1355 const char *path, *pr_path; 1356 size_t pr_pathlen; 1357 bool visible; 1358 1359 sx_assert(&shm_dict_lock, SA_LOCKED); 1360 kif->kf_type = KF_TYPE_SHM; 1361 kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode; 1362 kif->kf_un.kf_file.kf_file_size = shmfd->shm_size; 1363 if (shmfd->shm_path != NULL) { 1364 if (shmfd->shm_path != NULL) { 1365 path = shmfd->shm_path; 1366 pr_path = curthread->td_ucred->cr_prison->pr_path; 1367 if (strcmp(pr_path, "/") != 0) { 1368 /* Return the jail-rooted pathname. */ 1369 pr_pathlen = strlen(pr_path); 1370 visible = strncmp(path, pr_path, pr_pathlen) 1371 == 0 && path[pr_pathlen] == '/'; 1372 if (list && !visible) 1373 return (EPERM); 1374 if (visible) 1375 path += pr_pathlen; 1376 } 1377 strlcpy(kif->kf_path, path, sizeof(kif->kf_path)); 1378 } 1379 } 1380 return (0); 1381 } 1382 1383 static int 1384 shm_fill_kinfo(struct file *fp, struct kinfo_file *kif, 1385 struct filedesc *fdp __unused) 1386 { 1387 int res; 1388 1389 sx_slock(&shm_dict_lock); 1390 res = shm_fill_kinfo_locked(fp->f_data, kif, false); 1391 sx_sunlock(&shm_dict_lock); 1392 return (res); 1393 } 1394 1395 static int 1396 shm_add_seals(struct file *fp, int seals) 1397 { 1398 struct shmfd *shmfd; 1399 void *rl_cookie; 1400 vm_ooffset_t writemappings; 1401 int error, nseals; 1402 1403 error = 0; 1404 shmfd = fp->f_data; 1405 rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX, 1406 &shmfd->shm_mtx); 1407 1408 /* Even already-set seals should result in EPERM. */ 1409 if ((shmfd->shm_seals & F_SEAL_SEAL) != 0) { 1410 error = EPERM; 1411 goto out; 1412 } 1413 nseals = seals & ~shmfd->shm_seals; 1414 if ((nseals & F_SEAL_WRITE) != 0) { 1415 /* 1416 * The rangelock above prevents writable mappings from being 1417 * added after we've started applying seals. The RLOCK here 1418 * is to avoid torn reads on ILP32 arches as unmapping/reducing 1419 * writemappings will be done without a rangelock. 1420 */ 1421 VM_OBJECT_RLOCK(shmfd->shm_object); 1422 writemappings = shmfd->shm_object->un_pager.swp.writemappings; 1423 VM_OBJECT_RUNLOCK(shmfd->shm_object); 1424 /* kmappings are also writable */ 1425 if (writemappings > 0) { 1426 error = EBUSY; 1427 goto out; 1428 } 1429 } 1430 shmfd->shm_seals |= nseals; 1431 out: 1432 rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); 1433 return (error); 1434 } 1435 1436 static int 1437 shm_get_seals(struct file *fp, int *seals) 1438 { 1439 struct shmfd *shmfd; 1440 1441 shmfd = fp->f_data; 1442 *seals = shmfd->shm_seals; 1443 return (0); 1444 } 1445 1446 static int 1447 shm_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td) 1448 { 1449 void *rl_cookie; 1450 struct shmfd *shmfd; 1451 size_t size; 1452 int error; 1453 1454 /* This assumes that the caller already checked for overflow. */ 1455 error = 0; 1456 shmfd = fp->f_data; 1457 size = offset + len; 1458 1459 /* 1460 * Just grab the rangelock for the range that we may be attempting to 1461 * grow, rather than blocking read/write for regions we won't be 1462 * touching while this (potential) resize is in progress. Other 1463 * attempts to resize the shmfd will have to take a write lock from 0 to 1464 * OFF_MAX, so this being potentially beyond the current usable range of 1465 * the shmfd is not necessarily a concern. If other mechanisms are 1466 * added to grow a shmfd, this may need to be re-evaluated. 1467 */ 1468 rl_cookie = rangelock_wlock(&shmfd->shm_rl, offset, size, 1469 &shmfd->shm_mtx); 1470 if (size > shmfd->shm_size) { 1471 VM_OBJECT_WLOCK(shmfd->shm_object); 1472 error = shm_dotruncate_locked(shmfd, size, rl_cookie); 1473 VM_OBJECT_WUNLOCK(shmfd->shm_object); 1474 } 1475 rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); 1476 /* Translate to posix_fallocate(2) return value as needed. */ 1477 if (error == ENOMEM) 1478 error = ENOSPC; 1479 return (error); 1480 } 1481 1482 static int 1483 sysctl_posix_shm_list(SYSCTL_HANDLER_ARGS) 1484 { 1485 struct shm_mapping *shmm; 1486 struct sbuf sb; 1487 struct kinfo_file kif; 1488 u_long i; 1489 ssize_t curlen; 1490 int error, error2; 1491 1492 sbuf_new_for_sysctl(&sb, NULL, sizeof(struct kinfo_file) * 5, req); 1493 sbuf_clear_flags(&sb, SBUF_INCLUDENUL); 1494 curlen = 0; 1495 error = 0; 1496 sx_slock(&shm_dict_lock); 1497 for (i = 0; i < shm_hash + 1; i++) { 1498 LIST_FOREACH(shmm, &shm_dictionary[i], sm_link) { 1499 error = shm_fill_kinfo_locked(shmm->sm_shmfd, 1500 &kif, true); 1501 if (error == EPERM) 1502 continue; 1503 if (error != 0) 1504 break; 1505 pack_kinfo(&kif); 1506 if (req->oldptr != NULL && 1507 kif.kf_structsize + curlen > req->oldlen) 1508 break; 1509 error = sbuf_bcat(&sb, &kif, kif.kf_structsize) == 0 ? 1510 0 : ENOMEM; 1511 if (error != 0) 1512 break; 1513 curlen += kif.kf_structsize; 1514 } 1515 } 1516 sx_sunlock(&shm_dict_lock); 1517 error2 = sbuf_finish(&sb); 1518 sbuf_delete(&sb); 1519 return (error != 0 ? error : error2); 1520 } 1521 1522 SYSCTL_PROC(_kern_ipc, OID_AUTO, posix_shm_list, 1523 CTLFLAG_RD | CTLFLAG_MPSAFE | CTLTYPE_OPAQUE, 1524 NULL, 0, sysctl_posix_shm_list, "", 1525 "POSIX SHM list"); 1526 1527 int 1528 kern_shm_open(struct thread *td, const char *path, int flags, mode_t mode, 1529 struct filecaps *caps) 1530 { 1531 1532 return (kern_shm_open2(td, path, flags, mode, 0, caps, NULL)); 1533 } 1534 1535 /* 1536 * This version of the shm_open() interface leaves CLOEXEC behavior up to the 1537 * caller, and libc will enforce it for the traditional shm_open() call. This 1538 * allows other consumers, like memfd_create(), to opt-in for CLOEXEC. This 1539 * interface also includes a 'name' argument that is currently unused, but could 1540 * potentially be exported later via some interface for debugging purposes. 1541 * From the kernel's perspective, it is optional. Individual consumers like 1542 * memfd_create() may require it in order to be compatible with other systems 1543 * implementing the same function. 1544 */ 1545 int 1546 sys_shm_open2(struct thread *td, struct shm_open2_args *uap) 1547 { 1548 1549 return (kern_shm_open2(td, uap->path, uap->flags, uap->mode, 1550 uap->shmflags, NULL, uap->name)); 1551 } 1552