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