sys/kern/uipc_shm.c

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