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 rv = vm_page_grab_valid_unlocked(&m, obj, idx, 180 VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY | VM_ALLOC_NOCREAT); 181 if (rv == VM_PAGER_OK) 182 goto found; 183 184 /* 185 * Read I/O without either a corresponding resident page or swap 186 * page: use zero_region. This is intended to avoid instantiating 187 * pages on read from a sparse region. 188 */ 189 VM_OBJECT_WLOCK(obj); 190 m = vm_page_lookup(obj, idx); 191 if (uio->uio_rw == UIO_READ && m == NULL && 192 !vm_pager_has_page(obj, idx, NULL, NULL)) { 193 VM_OBJECT_WUNLOCK(obj); 194 return (uiomove(__DECONST(void *, zero_region), tlen, uio)); 195 } 196 197 /* 198 * Although the tmpfs vnode lock is held here, it is 199 * nonetheless safe to sleep waiting for a free page. The 200 * pageout daemon does not need to acquire the tmpfs vnode 201 * lock to page out tobj's pages because tobj is a OBJT_SWAP 202 * type object. 203 */ 204 rv = vm_page_grab_valid(&m, obj, idx, 205 VM_ALLOC_NORMAL | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY); 206 if (rv != VM_PAGER_OK) { 207 VM_OBJECT_WUNLOCK(obj); 208 printf("uiomove_object: vm_obj %p idx %jd pager error %d\n", 209 obj, idx, rv); 210 return (EIO); 211 } 212 VM_OBJECT_WUNLOCK(obj); 213 214 found: 215 error = uiomove_fromphys(&m, offset, tlen, uio); 216 if (uio->uio_rw == UIO_WRITE && error == 0) 217 vm_page_set_dirty(m); 218 vm_page_activate(m); 219 vm_page_sunbusy(m); 220 221 return (error); 222 } 223 224 int 225 uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio) 226 { 227 ssize_t resid; 228 size_t len; 229 int error; 230 231 error = 0; 232 while ((resid = uio->uio_resid) > 0) { 233 if (obj_size <= uio->uio_offset) 234 break; 235 len = MIN(obj_size - uio->uio_offset, resid); 236 if (len == 0) 237 break; 238 error = uiomove_object_page(obj, len, uio); 239 if (error != 0 || resid == uio->uio_resid) 240 break; 241 } 242 return (error); 243 } 244 245 static int 246 shm_seek(struct file *fp, off_t offset, int whence, struct thread *td) 247 { 248 struct shmfd *shmfd; 249 off_t foffset; 250 int error; 251 252 shmfd = fp->f_data; 253 foffset = foffset_lock(fp, 0); 254 error = 0; 255 switch (whence) { 256 case L_INCR: 257 if (foffset < 0 || 258 (offset > 0 && foffset > OFF_MAX - offset)) { 259 error = EOVERFLOW; 260 break; 261 } 262 offset += foffset; 263 break; 264 case L_XTND: 265 if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) { 266 error = EOVERFLOW; 267 break; 268 } 269 offset += shmfd->shm_size; 270 break; 271 case L_SET: 272 break; 273 default: 274 error = EINVAL; 275 } 276 if (error == 0) { 277 if (offset < 0 || offset > shmfd->shm_size) 278 error = EINVAL; 279 else 280 td->td_uretoff.tdu_off = offset; 281 } 282 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0); 283 return (error); 284 } 285 286 static int 287 shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred, 288 int flags, struct thread *td) 289 { 290 struct shmfd *shmfd; 291 void *rl_cookie; 292 int error; 293 294 shmfd = fp->f_data; 295 #ifdef MAC 296 error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd); 297 if (error) 298 return (error); 299 #endif 300 foffset_lock_uio(fp, uio, flags); 301 rl_cookie = rangelock_rlock(&shmfd->shm_rl, uio->uio_offset, 302 uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx); 303 error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio); 304 rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); 305 foffset_unlock_uio(fp, uio, flags); 306 return (error); 307 } 308 309 static int 310 shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred, 311 int flags, struct thread *td) 312 { 313 struct shmfd *shmfd; 314 void *rl_cookie; 315 int error; 316 317 shmfd = fp->f_data; 318 #ifdef MAC 319 error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd); 320 if (error) 321 return (error); 322 #endif 323 foffset_lock_uio(fp, uio, flags); 324 if ((flags & FOF_OFFSET) == 0) { 325 rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX, 326 &shmfd->shm_mtx); 327 } else { 328 rl_cookie = rangelock_wlock(&shmfd->shm_rl, uio->uio_offset, 329 uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx); 330 } 331 if ((shmfd->shm_seals & F_SEAL_WRITE) != 0) 332 error = EPERM; 333 else 334 error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio); 335 rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); 336 foffset_unlock_uio(fp, uio, flags); 337 return (error); 338 } 339 340 static int 341 shm_truncate(struct file *fp, off_t length, struct ucred *active_cred, 342 struct thread *td) 343 { 344 struct shmfd *shmfd; 345 #ifdef MAC 346 int error; 347 #endif 348 349 shmfd = fp->f_data; 350 #ifdef MAC 351 error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd); 352 if (error) 353 return (error); 354 #endif 355 return (shm_dotruncate(shmfd, length)); 356 } 357 358 int 359 shm_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred, 360 struct thread *td) 361 { 362 363 switch (com) { 364 case FIONBIO: 365 case FIOASYNC: 366 /* 367 * Allow fcntl(fd, F_SETFL, O_NONBLOCK) to work, 368 * just like it would on an unlinked regular file 369 */ 370 return (0); 371 default: 372 return (ENOTTY); 373 } 374 } 375 376 static int 377 shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred, 378 struct thread *td) 379 { 380 struct shmfd *shmfd; 381 #ifdef MAC 382 int error; 383 #endif 384 385 shmfd = fp->f_data; 386 387 #ifdef MAC 388 error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd); 389 if (error) 390 return (error); 391 #endif 392 393 /* 394 * Attempt to return sanish values for fstat() on a memory file 395 * descriptor. 396 */ 397 bzero(sb, sizeof(*sb)); 398 sb->st_blksize = PAGE_SIZE; 399 sb->st_size = shmfd->shm_size; 400 sb->st_blocks = howmany(sb->st_size, sb->st_blksize); 401 mtx_lock(&shm_timestamp_lock); 402 sb->st_atim = shmfd->shm_atime; 403 sb->st_ctim = shmfd->shm_ctime; 404 sb->st_mtim = shmfd->shm_mtime; 405 sb->st_birthtim = shmfd->shm_birthtime; 406 sb->st_mode = S_IFREG | shmfd->shm_mode; /* XXX */ 407 sb->st_uid = shmfd->shm_uid; 408 sb->st_gid = shmfd->shm_gid; 409 mtx_unlock(&shm_timestamp_lock); 410 sb->st_dev = shm_dev_ino; 411 sb->st_ino = shmfd->shm_ino; 412 sb->st_nlink = shmfd->shm_object->ref_count; 413 414 return (0); 415 } 416 417 static int 418 shm_close(struct file *fp, struct thread *td) 419 { 420 struct shmfd *shmfd; 421 422 shmfd = fp->f_data; 423 fp->f_data = NULL; 424 shm_drop(shmfd); 425 426 return (0); 427 } 428 429 static int 430 shm_copyin_path(struct thread *td, const char *userpath_in, char **path_out) { 431 int error; 432 char *path; 433 const char *pr_path; 434 size_t pr_pathlen; 435 436 path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK); 437 pr_path = td->td_ucred->cr_prison->pr_path; 438 439 /* Construct a full pathname for jailed callers. */ 440 pr_pathlen = strcmp(pr_path, "/") == 441 0 ? 0 : strlcpy(path, pr_path, MAXPATHLEN); 442 error = copyinstr(userpath_in, path + pr_pathlen, 443 MAXPATHLEN - pr_pathlen, NULL); 444 if (error != 0) 445 goto out; 446 447 #ifdef KTRACE 448 if (KTRPOINT(curthread, KTR_NAMEI)) 449 ktrnamei(path); 450 #endif 451 452 /* Require paths to start with a '/' character. */ 453 if (path[pr_pathlen] != '/') { 454 error = EINVAL; 455 goto out; 456 } 457 458 *path_out = path; 459 460 out: 461 if (error != 0) 462 free(path, M_SHMFD); 463 464 return (error); 465 } 466 467 static int 468 shm_dotruncate_locked(struct shmfd *shmfd, off_t length, void *rl_cookie) 469 { 470 vm_object_t object; 471 vm_page_t m; 472 vm_pindex_t idx, nobjsize; 473 vm_ooffset_t delta; 474 int base, rv; 475 476 KASSERT(length >= 0, ("shm_dotruncate: length < 0")); 477 object = shmfd->shm_object; 478 VM_OBJECT_ASSERT_WLOCKED(object); 479 rangelock_cookie_assert(rl_cookie, RA_WLOCKED); 480 if (length == shmfd->shm_size) 481 return (0); 482 nobjsize = OFF_TO_IDX(length + PAGE_MASK); 483 484 /* Are we shrinking? If so, trim the end. */ 485 if (length < shmfd->shm_size) { 486 if ((shmfd->shm_seals & F_SEAL_SHRINK) != 0) 487 return (EPERM); 488 489 /* 490 * Disallow any requests to shrink the size if this 491 * object is mapped into the kernel. 492 */ 493 if (shmfd->shm_kmappings > 0) 494 return (EBUSY); 495 496 /* 497 * Zero the truncated part of the last page. 498 */ 499 base = length & PAGE_MASK; 500 if (base != 0) { 501 idx = OFF_TO_IDX(length); 502 retry: 503 m = vm_page_grab(object, idx, VM_ALLOC_NOCREAT); 504 if (m != NULL) { 505 MPASS(vm_page_all_valid(m)); 506 } else if (vm_pager_has_page(object, idx, NULL, NULL)) { 507 m = vm_page_alloc(object, idx, 508 VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL); 509 if (m == NULL) 510 goto retry; 511 vm_object_pip_add(object, 1); 512 VM_OBJECT_WUNLOCK(object); 513 rv = vm_pager_get_pages(object, &m, 1, NULL, 514 NULL); 515 VM_OBJECT_WLOCK(object); 516 vm_object_pip_wakeup(object); 517 if (rv == VM_PAGER_OK) { 518 /* 519 * Since the page was not resident, 520 * and therefore not recently 521 * accessed, immediately enqueue it 522 * for asynchronous laundering. The 523 * current operation is not regarded 524 * as an access. 525 */ 526 vm_page_launder(m); 527 } else { 528 vm_page_free(m); 529 VM_OBJECT_WUNLOCK(object); 530 return (EIO); 531 } 532 } 533 if (m != NULL) { 534 pmap_zero_page_area(m, base, PAGE_SIZE - base); 535 KASSERT(vm_page_all_valid(m), 536 ("shm_dotruncate: page %p is invalid", m)); 537 vm_page_set_dirty(m); 538 vm_page_xunbusy(m); 539 } 540 } 541 delta = IDX_TO_OFF(object->size - nobjsize); 542 543 /* Toss in memory pages. */ 544 if (nobjsize < object->size) 545 vm_object_page_remove(object, nobjsize, object->size, 546 0); 547 548 /* Toss pages from swap. */ 549 if (object->type == OBJT_SWAP) 550 swap_pager_freespace(object, nobjsize, delta); 551 552 /* Free the swap accounted for shm */ 553 swap_release_by_cred(delta, object->cred); 554 object->charge -= delta; 555 } else { 556 if ((shmfd->shm_seals & F_SEAL_GROW) != 0) 557 return (EPERM); 558 559 /* Try to reserve additional swap space. */ 560 delta = IDX_TO_OFF(nobjsize - object->size); 561 if (!swap_reserve_by_cred(delta, object->cred)) 562 return (ENOMEM); 563 object->charge += delta; 564 } 565 shmfd->shm_size = length; 566 mtx_lock(&shm_timestamp_lock); 567 vfs_timestamp(&shmfd->shm_ctime); 568 shmfd->shm_mtime = shmfd->shm_ctime; 569 mtx_unlock(&shm_timestamp_lock); 570 object->size = nobjsize; 571 return (0); 572 } 573 574 int 575 shm_dotruncate(struct shmfd *shmfd, off_t length) 576 { 577 void *rl_cookie; 578 int error; 579 580 rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX, 581 &shmfd->shm_mtx); 582 VM_OBJECT_WLOCK(shmfd->shm_object); 583 error = shm_dotruncate_locked(shmfd, length, rl_cookie); 584 VM_OBJECT_WUNLOCK(shmfd->shm_object); 585 rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); 586 return (error); 587 } 588 589 /* 590 * shmfd object management including creation and reference counting 591 * routines. 592 */ 593 struct shmfd * 594 shm_alloc(struct ucred *ucred, mode_t mode) 595 { 596 struct shmfd *shmfd; 597 598 shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO); 599 shmfd->shm_size = 0; 600 shmfd->shm_uid = ucred->cr_uid; 601 shmfd->shm_gid = ucred->cr_gid; 602 shmfd->shm_mode = mode; 603 shmfd->shm_object = vm_pager_allocate(OBJT_SWAP, NULL, 604 shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred); 605 KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate")); 606 vfs_timestamp(&shmfd->shm_birthtime); 607 shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime = 608 shmfd->shm_birthtime; 609 shmfd->shm_ino = alloc_unr64(&shm_ino_unr); 610 refcount_init(&shmfd->shm_refs, 1); 611 mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF); 612 rangelock_init(&shmfd->shm_rl); 613 #ifdef MAC 614 mac_posixshm_init(shmfd); 615 mac_posixshm_create(ucred, shmfd); 616 #endif 617 618 return (shmfd); 619 } 620 621 struct shmfd * 622 shm_hold(struct shmfd *shmfd) 623 { 624 625 refcount_acquire(&shmfd->shm_refs); 626 return (shmfd); 627 } 628 629 void 630 shm_drop(struct shmfd *shmfd) 631 { 632 633 if (refcount_release(&shmfd->shm_refs)) { 634 #ifdef MAC 635 mac_posixshm_destroy(shmfd); 636 #endif 637 rangelock_destroy(&shmfd->shm_rl); 638 mtx_destroy(&shmfd->shm_mtx); 639 vm_object_deallocate(shmfd->shm_object); 640 free(shmfd, M_SHMFD); 641 } 642 } 643 644 /* 645 * Determine if the credentials have sufficient permissions for a 646 * specified combination of FREAD and FWRITE. 647 */ 648 int 649 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags) 650 { 651 accmode_t accmode; 652 int error; 653 654 accmode = 0; 655 if (flags & FREAD) 656 accmode |= VREAD; 657 if (flags & FWRITE) 658 accmode |= VWRITE; 659 mtx_lock(&shm_timestamp_lock); 660 error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid, 661 accmode, ucred, NULL); 662 mtx_unlock(&shm_timestamp_lock); 663 return (error); 664 } 665 666 /* 667 * Dictionary management. We maintain an in-kernel dictionary to map 668 * paths to shmfd objects. We use the FNV hash on the path to store 669 * the mappings in a hash table. 670 */ 671 static void 672 shm_init(void *arg) 673 { 674 675 mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF); 676 sx_init(&shm_dict_lock, "shm dictionary"); 677 shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash); 678 new_unrhdr64(&shm_ino_unr, 1); 679 shm_dev_ino = devfs_alloc_cdp_inode(); 680 KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized")); 681 } 682 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL); 683 684 static struct shmfd * 685 shm_lookup(char *path, Fnv32_t fnv) 686 { 687 struct shm_mapping *map; 688 689 LIST_FOREACH(map, SHM_HASH(fnv), sm_link) { 690 if (map->sm_fnv != fnv) 691 continue; 692 if (strcmp(map->sm_path, path) == 0) 693 return (map->sm_shmfd); 694 } 695 696 return (NULL); 697 } 698 699 static void 700 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd) 701 { 702 struct shm_mapping *map; 703 704 map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK); 705 map->sm_path = path; 706 map->sm_fnv = fnv; 707 map->sm_shmfd = shm_hold(shmfd); 708 shmfd->shm_path = path; 709 LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link); 710 } 711 712 static int 713 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred) 714 { 715 struct shm_mapping *map; 716 int error; 717 718 LIST_FOREACH(map, SHM_HASH(fnv), sm_link) { 719 if (map->sm_fnv != fnv) 720 continue; 721 if (strcmp(map->sm_path, path) == 0) { 722 #ifdef MAC 723 error = mac_posixshm_check_unlink(ucred, map->sm_shmfd); 724 if (error) 725 return (error); 726 #endif 727 error = shm_access(map->sm_shmfd, ucred, 728 FREAD | FWRITE); 729 if (error) 730 return (error); 731 map->sm_shmfd->shm_path = NULL; 732 LIST_REMOVE(map, sm_link); 733 shm_drop(map->sm_shmfd); 734 free(map->sm_path, M_SHMFD); 735 free(map, M_SHMFD); 736 return (0); 737 } 738 } 739 740 return (ENOENT); 741 } 742 743 int 744 kern_shm_open2(struct thread *td, const char *userpath, int flags, mode_t mode, 745 int shmflags, struct filecaps *fcaps, const char *name __unused) 746 { 747 struct filedesc *fdp; 748 struct shmfd *shmfd; 749 struct file *fp; 750 char *path; 751 void *rl_cookie; 752 Fnv32_t fnv; 753 mode_t cmode; 754 int error, fd, initial_seals; 755 756 if ((shmflags & ~SHM_ALLOW_SEALING) != 0) 757 return (EINVAL); 758 759 initial_seals = F_SEAL_SEAL; 760 if ((shmflags & SHM_ALLOW_SEALING) != 0) 761 initial_seals &= ~F_SEAL_SEAL; 762 763 #ifdef CAPABILITY_MODE 764 /* 765 * shm_open(2) is only allowed for anonymous objects. 766 */ 767 if (IN_CAPABILITY_MODE(td) && (userpath != SHM_ANON)) 768 return (ECAPMODE); 769 #endif 770 771 AUDIT_ARG_FFLAGS(flags); 772 AUDIT_ARG_MODE(mode); 773 774 if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR) 775 return (EINVAL); 776 777 if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0) 778 return (EINVAL); 779 780 /* 781 * Currently only F_SEAL_SEAL may be set when creating or opening shmfd. 782 * If the decision is made later to allow additional seals, care must be 783 * taken below to ensure that the seals are properly set if the shmfd 784 * already existed -- this currently assumes that only F_SEAL_SEAL can 785 * be set and doesn't take further precautions to ensure the validity of 786 * the seals being added with respect to current mappings. 787 */ 788 if ((initial_seals & ~F_SEAL_SEAL) != 0) 789 return (EINVAL); 790 791 fdp = td->td_proc->p_fd; 792 cmode = (mode & ~fdp->fd_cmask) & ACCESSPERMS; 793 794 /* 795 * shm_open(2) created shm should always have O_CLOEXEC set, as mandated 796 * by POSIX. We allow it to be unset here so that an in-kernel 797 * interface may be written as a thin layer around shm, optionally not 798 * setting CLOEXEC. For shm_open(2), O_CLOEXEC is set unconditionally 799 * in sys_shm_open() to keep this implementation compliant. 800 */ 801 error = falloc_caps(td, &fp, &fd, flags & O_CLOEXEC, fcaps); 802 if (error) 803 return (error); 804 805 /* A SHM_ANON path pointer creates an anonymous object. */ 806 if (userpath == SHM_ANON) { 807 /* A read-only anonymous object is pointless. */ 808 if ((flags & O_ACCMODE) == O_RDONLY) { 809 fdclose(td, fp, fd); 810 fdrop(fp, td); 811 return (EINVAL); 812 } 813 shmfd = shm_alloc(td->td_ucred, cmode); 814 shmfd->shm_seals = initial_seals; 815 } else { 816 error = shm_copyin_path(td, userpath, &path); 817 if (error != 0) { 818 fdclose(td, fp, fd); 819 fdrop(fp, td); 820 return (error); 821 } 822 823 AUDIT_ARG_UPATH1_CANON(path); 824 fnv = fnv_32_str(path, FNV1_32_INIT); 825 sx_xlock(&shm_dict_lock); 826 shmfd = shm_lookup(path, fnv); 827 if (shmfd == NULL) { 828 /* Object does not yet exist, create it if requested. */ 829 if (flags & O_CREAT) { 830 #ifdef MAC 831 error = mac_posixshm_check_create(td->td_ucred, 832 path); 833 if (error == 0) { 834 #endif 835 shmfd = shm_alloc(td->td_ucred, cmode); 836 shmfd->shm_seals = initial_seals; 837 shm_insert(path, fnv, shmfd); 838 #ifdef MAC 839 } 840 #endif 841 } else { 842 free(path, M_SHMFD); 843 error = ENOENT; 844 } 845 } else { 846 rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX, 847 &shmfd->shm_mtx); 848 849 /* 850 * kern_shm_open() likely shouldn't ever error out on 851 * trying to set a seal that already exists, unlike 852 * F_ADD_SEALS. This would break terribly as 853 * shm_open(2) actually sets F_SEAL_SEAL to maintain 854 * historical behavior where the underlying file could 855 * not be sealed. 856 */ 857 initial_seals &= ~shmfd->shm_seals; 858 859 /* 860 * Object already exists, obtain a new 861 * reference if requested and permitted. 862 */ 863 free(path, M_SHMFD); 864 865 /* 866 * initial_seals can't set additional seals if we've 867 * already been set F_SEAL_SEAL. If F_SEAL_SEAL is set, 868 * then we've already removed that one from 869 * initial_seals. This is currently redundant as we 870 * only allow setting F_SEAL_SEAL at creation time, but 871 * it's cheap to check and decreases the effort required 872 * to allow additional seals. 873 */ 874 if ((shmfd->shm_seals & F_SEAL_SEAL) != 0 && 875 initial_seals != 0) 876 error = EPERM; 877 else if ((flags & (O_CREAT | O_EXCL)) == 878 (O_CREAT | O_EXCL)) 879 error = EEXIST; 880 else { 881 #ifdef MAC 882 error = mac_posixshm_check_open(td->td_ucred, 883 shmfd, FFLAGS(flags & O_ACCMODE)); 884 if (error == 0) 885 #endif 886 error = shm_access(shmfd, td->td_ucred, 887 FFLAGS(flags & O_ACCMODE)); 888 } 889 890 /* 891 * Truncate the file back to zero length if 892 * O_TRUNC was specified and the object was 893 * opened with read/write. 894 */ 895 if (error == 0 && 896 (flags & (O_ACCMODE | O_TRUNC)) == 897 (O_RDWR | O_TRUNC)) { 898 VM_OBJECT_WLOCK(shmfd->shm_object); 899 #ifdef MAC 900 error = mac_posixshm_check_truncate( 901 td->td_ucred, fp->f_cred, shmfd); 902 if (error == 0) 903 #endif 904 error = shm_dotruncate_locked(shmfd, 0, 905 rl_cookie); 906 VM_OBJECT_WUNLOCK(shmfd->shm_object); 907 } 908 if (error == 0) { 909 /* 910 * Currently we only allow F_SEAL_SEAL to be 911 * set initially. As noted above, this would 912 * need to be reworked should that change. 913 */ 914 shmfd->shm_seals |= initial_seals; 915 shm_hold(shmfd); 916 } 917 rangelock_unlock(&shmfd->shm_rl, rl_cookie, 918 &shmfd->shm_mtx); 919 } 920 sx_xunlock(&shm_dict_lock); 921 922 if (error) { 923 fdclose(td, fp, fd); 924 fdrop(fp, td); 925 return (error); 926 } 927 } 928 929 finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops); 930 931 td->td_retval[0] = fd; 932 fdrop(fp, td); 933 934 return (0); 935 } 936 937 /* System calls. */ 938 #ifdef COMPAT_FREEBSD12 939 int 940 freebsd12_shm_open(struct thread *td, struct freebsd12_shm_open_args *uap) 941 { 942 943 return (kern_shm_open(td, uap->path, uap->flags | O_CLOEXEC, 944 uap->mode, NULL)); 945 } 946 #endif 947 948 int 949 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap) 950 { 951 char *path; 952 Fnv32_t fnv; 953 int error; 954 955 error = shm_copyin_path(td, uap->path, &path); 956 if (error != 0) 957 return (error); 958 959 AUDIT_ARG_UPATH1_CANON(path); 960 fnv = fnv_32_str(path, FNV1_32_INIT); 961 sx_xlock(&shm_dict_lock); 962 error = shm_remove(path, fnv, td->td_ucred); 963 sx_xunlock(&shm_dict_lock); 964 free(path, M_SHMFD); 965 966 return (error); 967 } 968 969 int 970 sys_shm_rename(struct thread *td, struct shm_rename_args *uap) 971 { 972 char *path_from = NULL, *path_to = NULL; 973 Fnv32_t fnv_from, fnv_to; 974 struct shmfd *fd_from; 975 struct shmfd *fd_to; 976 int error; 977 int flags; 978 979 flags = uap->flags; 980 AUDIT_ARG_FFLAGS(flags); 981 982 /* 983 * Make sure the user passed only valid flags. 984 * If you add a new flag, please add a new term here. 985 */ 986 if ((flags & ~( 987 SHM_RENAME_NOREPLACE | 988 SHM_RENAME_EXCHANGE 989 )) != 0) { 990 error = EINVAL; 991 goto out; 992 } 993 994 /* 995 * EXCHANGE and NOREPLACE don't quite make sense together. Let's 996 * force the user to choose one or the other. 997 */ 998 if ((flags & SHM_RENAME_NOREPLACE) != 0 && 999 (flags & SHM_RENAME_EXCHANGE) != 0) { 1000 error = EINVAL; 1001 goto out; 1002 } 1003 1004 /* Renaming to or from anonymous makes no sense */ 1005 if (uap->path_from == SHM_ANON || uap->path_to == SHM_ANON) { 1006 error = EINVAL; 1007 goto out; 1008 } 1009 1010 error = shm_copyin_path(td, uap->path_from, &path_from); 1011 if (error != 0) 1012 goto out; 1013 1014 error = shm_copyin_path(td, uap->path_to, &path_to); 1015 if (error != 0) 1016 goto out; 1017 1018 AUDIT_ARG_UPATH1_CANON(path_from); 1019 AUDIT_ARG_UPATH2_CANON(path_to); 1020 1021 /* Rename with from/to equal is a no-op */ 1022 if (strcmp(path_from, path_to) == 0) 1023 goto out; 1024 1025 fnv_from = fnv_32_str(path_from, FNV1_32_INIT); 1026 fnv_to = fnv_32_str(path_to, FNV1_32_INIT); 1027 1028 sx_xlock(&shm_dict_lock); 1029 1030 fd_from = shm_lookup(path_from, fnv_from); 1031 if (fd_from == NULL) { 1032 error = ENOENT; 1033 goto out_locked; 1034 } 1035 1036 fd_to = shm_lookup(path_to, fnv_to); 1037 if ((flags & SHM_RENAME_NOREPLACE) != 0 && fd_to != NULL) { 1038 error = EEXIST; 1039 goto out_locked; 1040 } 1041 1042 /* 1043 * Unconditionally prevents shm_remove from invalidating the 'from' 1044 * shm's state. 1045 */ 1046 shm_hold(fd_from); 1047 error = shm_remove(path_from, fnv_from, td->td_ucred); 1048 1049 /* 1050 * One of my assumptions failed if ENOENT (e.g. locking didn't 1051 * protect us) 1052 */ 1053 KASSERT(error != ENOENT, ("Our shm disappeared during shm_rename: %s", 1054 path_from)); 1055 if (error != 0) { 1056 shm_drop(fd_from); 1057 goto out_locked; 1058 } 1059 1060 /* 1061 * If we are exchanging, we need to ensure the shm_remove below 1062 * doesn't invalidate the dest shm's state. 1063 */ 1064 if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) 1065 shm_hold(fd_to); 1066 1067 /* 1068 * NOTE: if path_to is not already in the hash, c'est la vie; 1069 * it simply means we have nothing already at path_to to unlink. 1070 * That is the ENOENT case. 1071 * 1072 * If we somehow don't have access to unlink this guy, but 1073 * did for the shm at path_from, then relink the shm to path_from 1074 * and abort with EACCES. 1075 * 1076 * All other errors: that is weird; let's relink and abort the 1077 * operation. 1078 */ 1079 error = shm_remove(path_to, fnv_to, td->td_ucred); 1080 if (error != 0 && error != ENOENT) { 1081 shm_insert(path_from, fnv_from, fd_from); 1082 shm_drop(fd_from); 1083 /* Don't free path_from now, since the hash references it */ 1084 path_from = NULL; 1085 goto out_locked; 1086 } 1087 1088 error = 0; 1089 1090 shm_insert(path_to, fnv_to, fd_from); 1091 1092 /* Don't free path_to now, since the hash references it */ 1093 path_to = NULL; 1094 1095 /* We kept a ref when we removed, and incremented again in insert */ 1096 shm_drop(fd_from); 1097 KASSERT(fd_from->shm_refs > 0, ("Expected >0 refs; got: %d\n", 1098 fd_from->shm_refs)); 1099 1100 if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) { 1101 shm_insert(path_from, fnv_from, fd_to); 1102 path_from = NULL; 1103 shm_drop(fd_to); 1104 KASSERT(fd_to->shm_refs > 0, ("Expected >0 refs; got: %d\n", 1105 fd_to->shm_refs)); 1106 } 1107 1108 out_locked: 1109 sx_xunlock(&shm_dict_lock); 1110 1111 out: 1112 free(path_from, M_SHMFD); 1113 free(path_to, M_SHMFD); 1114 return (error); 1115 } 1116 1117 int 1118 shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize, 1119 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, 1120 vm_ooffset_t foff, struct thread *td) 1121 { 1122 struct shmfd *shmfd; 1123 vm_prot_t maxprot; 1124 int error; 1125 bool writecnt; 1126 void *rl_cookie; 1127 1128 shmfd = fp->f_data; 1129 maxprot = VM_PROT_NONE; 1130 1131 rl_cookie = rangelock_rlock(&shmfd->shm_rl, 0, objsize, 1132 &shmfd->shm_mtx); 1133 /* FREAD should always be set. */ 1134 if ((fp->f_flag & FREAD) != 0) 1135 maxprot |= VM_PROT_EXECUTE | VM_PROT_READ; 1136 1137 /* 1138 * If FWRITE's set, we can allow VM_PROT_WRITE unless it's a shared 1139 * mapping with a write seal applied. Private mappings are always 1140 * writeable. 1141 */ 1142 if ((flags & MAP_SHARED) == 0) { 1143 cap_maxprot |= VM_PROT_WRITE; 1144 maxprot |= VM_PROT_WRITE; 1145 writecnt = false; 1146 } else { 1147 if ((fp->f_flag & FWRITE) != 0 && 1148 (shmfd->shm_seals & F_SEAL_WRITE) == 0) 1149 maxprot |= VM_PROT_WRITE; 1150 writecnt = (prot & VM_PROT_WRITE) != 0; 1151 if (writecnt && (shmfd->shm_seals & F_SEAL_WRITE) != 0) { 1152 error = EPERM; 1153 goto out; 1154 } 1155 1156 /* Don't permit shared writable mappings on read-only descriptors. */ 1157 if (writecnt && (maxprot & VM_PROT_WRITE) == 0) { 1158 error = EACCES; 1159 goto out; 1160 } 1161 } 1162 maxprot &= cap_maxprot; 1163 1164 /* See comment in vn_mmap(). */ 1165 if ( 1166 #ifdef _LP64 1167 objsize > OFF_MAX || 1168 #endif 1169 foff < 0 || foff > OFF_MAX - objsize) { 1170 error = EINVAL; 1171 goto out; 1172 } 1173 1174 #ifdef MAC 1175 error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags); 1176 if (error != 0) 1177 goto out; 1178 #endif 1179 1180 mtx_lock(&shm_timestamp_lock); 1181 vfs_timestamp(&shmfd->shm_atime); 1182 mtx_unlock(&shm_timestamp_lock); 1183 vm_object_reference(shmfd->shm_object); 1184 1185 if (writecnt) 1186 vm_pager_update_writecount(shmfd->shm_object, 0, objsize); 1187 error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags, 1188 shmfd->shm_object, foff, writecnt, td); 1189 if (error != 0) { 1190 if (writecnt) 1191 vm_pager_release_writecount(shmfd->shm_object, 0, 1192 objsize); 1193 vm_object_deallocate(shmfd->shm_object); 1194 } 1195 out: 1196 rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); 1197 return (error); 1198 } 1199 1200 static int 1201 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, 1202 struct thread *td) 1203 { 1204 struct shmfd *shmfd; 1205 int error; 1206 1207 error = 0; 1208 shmfd = fp->f_data; 1209 mtx_lock(&shm_timestamp_lock); 1210 /* 1211 * SUSv4 says that x bits of permission need not be affected. 1212 * Be consistent with our shm_open there. 1213 */ 1214 #ifdef MAC 1215 error = mac_posixshm_check_setmode(active_cred, shmfd, mode); 1216 if (error != 0) 1217 goto out; 1218 #endif 1219 error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, 1220 shmfd->shm_gid, VADMIN, active_cred, NULL); 1221 if (error != 0) 1222 goto out; 1223 shmfd->shm_mode = mode & ACCESSPERMS; 1224 out: 1225 mtx_unlock(&shm_timestamp_lock); 1226 return (error); 1227 } 1228 1229 static int 1230 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred, 1231 struct thread *td) 1232 { 1233 struct shmfd *shmfd; 1234 int error; 1235 1236 error = 0; 1237 shmfd = fp->f_data; 1238 mtx_lock(&shm_timestamp_lock); 1239 #ifdef MAC 1240 error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid); 1241 if (error != 0) 1242 goto out; 1243 #endif 1244 if (uid == (uid_t)-1) 1245 uid = shmfd->shm_uid; 1246 if (gid == (gid_t)-1) 1247 gid = shmfd->shm_gid; 1248 if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) || 1249 (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) && 1250 (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN))) 1251 goto out; 1252 shmfd->shm_uid = uid; 1253 shmfd->shm_gid = gid; 1254 out: 1255 mtx_unlock(&shm_timestamp_lock); 1256 return (error); 1257 } 1258 1259 /* 1260 * Helper routines to allow the backing object of a shared memory file 1261 * descriptor to be mapped in the kernel. 1262 */ 1263 int 1264 shm_map(struct file *fp, size_t size, off_t offset, void **memp) 1265 { 1266 struct shmfd *shmfd; 1267 vm_offset_t kva, ofs; 1268 vm_object_t obj; 1269 int rv; 1270 1271 if (fp->f_type != DTYPE_SHM) 1272 return (EINVAL); 1273 shmfd = fp->f_data; 1274 obj = shmfd->shm_object; 1275 VM_OBJECT_WLOCK(obj); 1276 /* 1277 * XXXRW: This validation is probably insufficient, and subject to 1278 * sign errors. It should be fixed. 1279 */ 1280 if (offset >= shmfd->shm_size || 1281 offset + size > round_page(shmfd->shm_size)) { 1282 VM_OBJECT_WUNLOCK(obj); 1283 return (EINVAL); 1284 } 1285 1286 shmfd->shm_kmappings++; 1287 vm_object_reference_locked(obj); 1288 VM_OBJECT_WUNLOCK(obj); 1289 1290 /* Map the object into the kernel_map and wire it. */ 1291 kva = vm_map_min(kernel_map); 1292 ofs = offset & PAGE_MASK; 1293 offset = trunc_page(offset); 1294 size = round_page(size + ofs); 1295 rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0, 1296 VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE, 1297 VM_PROT_READ | VM_PROT_WRITE, 0); 1298 if (rv == KERN_SUCCESS) { 1299 rv = vm_map_wire(kernel_map, kva, kva + size, 1300 VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES); 1301 if (rv == KERN_SUCCESS) { 1302 *memp = (void *)(kva + ofs); 1303 return (0); 1304 } 1305 vm_map_remove(kernel_map, kva, kva + size); 1306 } else 1307 vm_object_deallocate(obj); 1308 1309 /* On failure, drop our mapping reference. */ 1310 VM_OBJECT_WLOCK(obj); 1311 shmfd->shm_kmappings--; 1312 VM_OBJECT_WUNLOCK(obj); 1313 1314 return (vm_mmap_to_errno(rv)); 1315 } 1316 1317 /* 1318 * We require the caller to unmap the entire entry. This allows us to 1319 * safely decrement shm_kmappings when a mapping is removed. 1320 */ 1321 int 1322 shm_unmap(struct file *fp, void *mem, size_t size) 1323 { 1324 struct shmfd *shmfd; 1325 vm_map_entry_t entry; 1326 vm_offset_t kva, ofs; 1327 vm_object_t obj; 1328 vm_pindex_t pindex; 1329 vm_prot_t prot; 1330 boolean_t wired; 1331 vm_map_t map; 1332 int rv; 1333 1334 if (fp->f_type != DTYPE_SHM) 1335 return (EINVAL); 1336 shmfd = fp->f_data; 1337 kva = (vm_offset_t)mem; 1338 ofs = kva & PAGE_MASK; 1339 kva = trunc_page(kva); 1340 size = round_page(size + ofs); 1341 map = kernel_map; 1342 rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry, 1343 &obj, &pindex, &prot, &wired); 1344 if (rv != KERN_SUCCESS) 1345 return (EINVAL); 1346 if (entry->start != kva || entry->end != kva + size) { 1347 vm_map_lookup_done(map, entry); 1348 return (EINVAL); 1349 } 1350 vm_map_lookup_done(map, entry); 1351 if (obj != shmfd->shm_object) 1352 return (EINVAL); 1353 vm_map_remove(map, kva, kva + size); 1354 VM_OBJECT_WLOCK(obj); 1355 KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped")); 1356 shmfd->shm_kmappings--; 1357 VM_OBJECT_WUNLOCK(obj); 1358 return (0); 1359 } 1360 1361 static int 1362 shm_fill_kinfo_locked(struct shmfd *shmfd, struct kinfo_file *kif, bool list) 1363 { 1364 const char *path, *pr_path; 1365 size_t pr_pathlen; 1366 bool visible; 1367 1368 sx_assert(&shm_dict_lock, SA_LOCKED); 1369 kif->kf_type = KF_TYPE_SHM; 1370 kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode; 1371 kif->kf_un.kf_file.kf_file_size = shmfd->shm_size; 1372 if (shmfd->shm_path != NULL) { 1373 if (shmfd->shm_path != NULL) { 1374 path = shmfd->shm_path; 1375 pr_path = curthread->td_ucred->cr_prison->pr_path; 1376 if (strcmp(pr_path, "/") != 0) { 1377 /* Return the jail-rooted pathname. */ 1378 pr_pathlen = strlen(pr_path); 1379 visible = strncmp(path, pr_path, pr_pathlen) 1380 == 0 && path[pr_pathlen] == '/'; 1381 if (list && !visible) 1382 return (EPERM); 1383 if (visible) 1384 path += pr_pathlen; 1385 } 1386 strlcpy(kif->kf_path, path, sizeof(kif->kf_path)); 1387 } 1388 } 1389 return (0); 1390 } 1391 1392 static int 1393 shm_fill_kinfo(struct file *fp, struct kinfo_file *kif, 1394 struct filedesc *fdp __unused) 1395 { 1396 int res; 1397 1398 sx_slock(&shm_dict_lock); 1399 res = shm_fill_kinfo_locked(fp->f_data, kif, false); 1400 sx_sunlock(&shm_dict_lock); 1401 return (res); 1402 } 1403 1404 static int 1405 shm_add_seals(struct file *fp, int seals) 1406 { 1407 struct shmfd *shmfd; 1408 void *rl_cookie; 1409 vm_ooffset_t writemappings; 1410 int error, nseals; 1411 1412 error = 0; 1413 shmfd = fp->f_data; 1414 rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX, 1415 &shmfd->shm_mtx); 1416 1417 /* Even already-set seals should result in EPERM. */ 1418 if ((shmfd->shm_seals & F_SEAL_SEAL) != 0) { 1419 error = EPERM; 1420 goto out; 1421 } 1422 nseals = seals & ~shmfd->shm_seals; 1423 if ((nseals & F_SEAL_WRITE) != 0) { 1424 /* 1425 * The rangelock above prevents writable mappings from being 1426 * added after we've started applying seals. The RLOCK here 1427 * is to avoid torn reads on ILP32 arches as unmapping/reducing 1428 * writemappings will be done without a rangelock. 1429 */ 1430 VM_OBJECT_RLOCK(shmfd->shm_object); 1431 writemappings = shmfd->shm_object->un_pager.swp.writemappings; 1432 VM_OBJECT_RUNLOCK(shmfd->shm_object); 1433 /* kmappings are also writable */ 1434 if (writemappings > 0) { 1435 error = EBUSY; 1436 goto out; 1437 } 1438 } 1439 shmfd->shm_seals |= nseals; 1440 out: 1441 rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); 1442 return (error); 1443 } 1444 1445 static int 1446 shm_get_seals(struct file *fp, int *seals) 1447 { 1448 struct shmfd *shmfd; 1449 1450 shmfd = fp->f_data; 1451 *seals = shmfd->shm_seals; 1452 return (0); 1453 } 1454 1455 static int 1456 shm_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td) 1457 { 1458 void *rl_cookie; 1459 struct shmfd *shmfd; 1460 size_t size; 1461 int error; 1462 1463 /* This assumes that the caller already checked for overflow. */ 1464 error = 0; 1465 shmfd = fp->f_data; 1466 size = offset + len; 1467 1468 /* 1469 * Just grab the rangelock for the range that we may be attempting to 1470 * grow, rather than blocking read/write for regions we won't be 1471 * touching while this (potential) resize is in progress. Other 1472 * attempts to resize the shmfd will have to take a write lock from 0 to 1473 * OFF_MAX, so this being potentially beyond the current usable range of 1474 * the shmfd is not necessarily a concern. If other mechanisms are 1475 * added to grow a shmfd, this may need to be re-evaluated. 1476 */ 1477 rl_cookie = rangelock_wlock(&shmfd->shm_rl, offset, size, 1478 &shmfd->shm_mtx); 1479 if (size > shmfd->shm_size) { 1480 VM_OBJECT_WLOCK(shmfd->shm_object); 1481 error = shm_dotruncate_locked(shmfd, size, rl_cookie); 1482 VM_OBJECT_WUNLOCK(shmfd->shm_object); 1483 } 1484 rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); 1485 /* Translate to posix_fallocate(2) return value as needed. */ 1486 if (error == ENOMEM) 1487 error = ENOSPC; 1488 return (error); 1489 } 1490 1491 static int 1492 sysctl_posix_shm_list(SYSCTL_HANDLER_ARGS) 1493 { 1494 struct shm_mapping *shmm; 1495 struct sbuf sb; 1496 struct kinfo_file kif; 1497 u_long i; 1498 ssize_t curlen; 1499 int error, error2; 1500 1501 sbuf_new_for_sysctl(&sb, NULL, sizeof(struct kinfo_file) * 5, req); 1502 sbuf_clear_flags(&sb, SBUF_INCLUDENUL); 1503 curlen = 0; 1504 error = 0; 1505 sx_slock(&shm_dict_lock); 1506 for (i = 0; i < shm_hash + 1; i++) { 1507 LIST_FOREACH(shmm, &shm_dictionary[i], sm_link) { 1508 error = shm_fill_kinfo_locked(shmm->sm_shmfd, 1509 &kif, true); 1510 if (error == EPERM) 1511 continue; 1512 if (error != 0) 1513 break; 1514 pack_kinfo(&kif); 1515 if (req->oldptr != NULL && 1516 kif.kf_structsize + curlen > req->oldlen) 1517 break; 1518 error = sbuf_bcat(&sb, &kif, kif.kf_structsize) == 0 ? 1519 0 : ENOMEM; 1520 if (error != 0) 1521 break; 1522 curlen += kif.kf_structsize; 1523 } 1524 } 1525 sx_sunlock(&shm_dict_lock); 1526 error2 = sbuf_finish(&sb); 1527 sbuf_delete(&sb); 1528 return (error != 0 ? error : error2); 1529 } 1530 1531 SYSCTL_PROC(_kern_ipc, OID_AUTO, posix_shm_list, 1532 CTLFLAG_RD | CTLFLAG_MPSAFE | CTLTYPE_OPAQUE, 1533 NULL, 0, sysctl_posix_shm_list, "", 1534 "POSIX SHM list"); 1535 1536 int 1537 kern_shm_open(struct thread *td, const char *path, int flags, mode_t mode, 1538 struct filecaps *caps) 1539 { 1540 1541 return (kern_shm_open2(td, path, flags, mode, 0, caps, NULL)); 1542 } 1543 1544 /* 1545 * This version of the shm_open() interface leaves CLOEXEC behavior up to the 1546 * caller, and libc will enforce it for the traditional shm_open() call. This 1547 * allows other consumers, like memfd_create(), to opt-in for CLOEXEC. This 1548 * interface also includes a 'name' argument that is currently unused, but could 1549 * potentially be exported later via some interface for debugging purposes. 1550 * From the kernel's perspective, it is optional. Individual consumers like 1551 * memfd_create() may require it in order to be compatible with other systems 1552 * implementing the same function. 1553 */ 1554 int 1555 sys_shm_open2(struct thread *td, struct shm_open2_args *uap) 1556 { 1557 1558 return (kern_shm_open2(td, uap->path, uap->flags, uap->mode, 1559 uap->shmflags, NULL, uap->name)); 1560 } 1561