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) and shm_unlink(2).  While most of the implementation is
  37  * here, vm_mmap.c contains mapping logic changes.
  38  *
  39  * posixshmcontrol(1) allows users to inspect the state of the memory
  40  * objects.  Per-uid swap resource limit controls total amount of
  41  * memory that user can consume for anonymous objects, including
  42  * shared.
  43  */
  44
  45 #include <sys/cdefs.h>
  46 __FBSDID("$FreeBSD$");
  47
  48 #include "opt_capsicum.h"
  49 #include "opt_ktrace.h"
  50
  51 #include <sys/param.h>
  52 #include <sys/capsicum.h>
  53 #include <sys/conf.h>
  54 #include <sys/fcntl.h>
  55 #include <sys/file.h>
  56 #include <sys/filedesc.h>
  57 #include <sys/filio.h>
  58 #include <sys/fnv_hash.h>
  59 #include <sys/kernel.h>
  60 #include <sys/limits.h>
  61 #include <sys/uio.h>
  62 #include <sys/signal.h>
  63 #include <sys/jail.h>
  64 #include <sys/ktrace.h>
  65 #include <sys/lock.h>
  66 #include <sys/malloc.h>
  67 #include <sys/mman.h>
  68 #include <sys/mutex.h>
  69 #include <sys/priv.h>
  70 #include <sys/proc.h>
  71 #include <sys/refcount.h>
  72 #include <sys/resourcevar.h>
  73 #include <sys/rwlock.h>
  74 #include <sys/sbuf.h>
  75 #include <sys/stat.h>
  76 #include <sys/syscallsubr.h>
  77 #include <sys/sysctl.h>
  78 #include <sys/sysproto.h>
  79 #include <sys/systm.h>
  80 #include <sys/sx.h>
  81 #include <sys/time.h>
  82 #include <sys/vnode.h>
  83 #include <sys/unistd.h>
  84 #include <sys/user.h>
  85
  86 #include <security/audit/audit.h>
  87 #include <security/mac/mac_framework.h>
  88
  89 #include <vm/vm.h>
  90 #include <vm/vm_param.h>
  91 #include <vm/pmap.h>
  92 #include <vm/vm_extern.h>
  93 #include <vm/vm_map.h>
  94 #include <vm/vm_kern.h>
  95 #include <vm/vm_object.h>
  96 #include <vm/vm_page.h>
  97 #include <vm/vm_pageout.h>
  98 #include <vm/vm_pager.h>
  99 #include <vm/swap_pager.h>
 100
 101 struct shm_mapping {
 102         char            *sm_path;
 103         Fnv32_t         sm_fnv;
 104         struct shmfd    *sm_shmfd;
 105         LIST_ENTRY(shm_mapping) sm_link;
 106 };
 107
 108 static MALLOC_DEFINE(M_SHMFD, "shmfd", "shared memory file descriptor");
 109 static LIST_HEAD(, shm_mapping) *shm_dictionary;
 110 static struct sx shm_dict_lock;
 111 static struct mtx shm_timestamp_lock;
 112 static u_long shm_hash;
 113 static struct unrhdr64 shm_ino_unr;
 114 static dev_t shm_dev_ino;
 115
 116 #define SHM_HASH(fnv)   (&shm_dictionary[(fnv) & shm_hash])
 117
 118 static void     shm_init(void *arg);
 119 static void     shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd);
 120 static struct shmfd *shm_lookup(char *path, Fnv32_t fnv);
 121 static int      shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred);
 122
 123 static fo_rdwr_t        shm_read;
 124 static fo_rdwr_t        shm_write;
 125 static fo_truncate_t    shm_truncate;
 126 static fo_ioctl_t       shm_ioctl;
 127 static fo_stat_t        shm_stat;
 128 static fo_close_t       shm_close;
 129 static fo_chmod_t       shm_chmod;
 130 static fo_chown_t       shm_chown;
 131 static fo_seek_t        shm_seek;
 132 static fo_fill_kinfo_t  shm_fill_kinfo;
 133 static fo_mmap_t        shm_mmap;
 134
 135 /* File descriptor operations. */
 136 struct fileops shm_ops = {
 137         .fo_read = shm_read,
 138         .fo_write = shm_write,
 139         .fo_truncate = shm_truncate,
 140         .fo_ioctl = shm_ioctl,
 141         .fo_poll = invfo_poll,
 142         .fo_kqfilter = invfo_kqfilter,
 143         .fo_stat = shm_stat,
 144         .fo_close = shm_close,
 145         .fo_chmod = shm_chmod,
 146         .fo_chown = shm_chown,
 147         .fo_sendfile = vn_sendfile,
 148         .fo_seek = shm_seek,
 149         .fo_fill_kinfo = shm_fill_kinfo,
 150         .fo_mmap = shm_mmap,
 151         .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
 152 };
 153
 154 FEATURE(posix_shm, "POSIX shared memory");
 155
 156 static int
 157 uiomove_object_page(vm_object_t obj, size_t len, struct uio *uio)
 158 {
 159         vm_page_t m;
 160         vm_pindex_t idx;
 161         size_t tlen;
 162         int error, offset, rv;
 163
 164         idx = OFF_TO_IDX(uio->uio_offset);
 165         offset = uio->uio_offset & PAGE_MASK;
 166         tlen = MIN(PAGE_SIZE - offset, len);
 167
 168         VM_OBJECT_WLOCK(obj);
 169
 170         /*
 171          * Read I/O without either a corresponding resident page or swap
 172          * page: use zero_region.  This is intended to avoid instantiating
 173          * pages on read from a sparse region.
 174          */
 175         if (uio->uio_rw == UIO_READ && vm_page_lookup(obj, idx) == NULL &&
 176             !vm_pager_has_page(obj, idx, NULL, NULL)) {
 177                 VM_OBJECT_WUNLOCK(obj);
 178                 return (uiomove(__DECONST(void *, zero_region), tlen, uio));
 179         }
 180
 181         /*
 182          * Parallel reads of the page content from disk are prevented
 183          * by exclusive busy.
 184          *
 185          * Although the tmpfs vnode lock is held here, it is
 186          * nonetheless safe to sleep waiting for a free page.  The
 187          * pageout daemon does not need to acquire the tmpfs vnode
 188          * lock to page out tobj's pages because tobj is a OBJT_SWAP
 189          * type object.
 190          */
 191         m = vm_page_grab(obj, idx, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY);
 192         if (m->valid != VM_PAGE_BITS_ALL) {
 193                 vm_page_xbusy(m);
 194                 if (vm_pager_has_page(obj, idx, NULL, NULL)) {
 195                         rv = vm_pager_get_pages(obj, &m, 1, NULL, NULL);
 196                         if (rv != VM_PAGER_OK) {
 197                                 printf(
 198             "uiomove_object: vm_obj %p idx %jd valid %x pager error %d\n",
 199                                     obj, idx, m->valid, rv);
 200                                 vm_page_lock(m);
 201                                 vm_page_free(m);
 202                                 vm_page_unlock(m);
 203                                 VM_OBJECT_WUNLOCK(obj);
 204                                 return (EIO);
 205                         }
 206                 } else
 207                         vm_page_zero_invalid(m, TRUE);
 208                 vm_page_xunbusy(m);
 209         }
 210         vm_page_lock(m);
 211         vm_page_hold(m);
 212         if (vm_page_active(m))
 213                 vm_page_reference(m);
 214         else
 215                 vm_page_activate(m);
 216         vm_page_unlock(m);
 217         VM_OBJECT_WUNLOCK(obj);
 218         error = uiomove_fromphys(&m, offset, tlen, uio);
 219         if (uio->uio_rw == UIO_WRITE && error == 0) {
 220                 VM_OBJECT_WLOCK(obj);
 221                 vm_page_dirty(m);
 222                 vm_pager_page_unswapped(m);
 223                 VM_OBJECT_WUNLOCK(obj);
 224         }
 225         vm_page_lock(m);
 226         vm_page_unhold(m);
 227         vm_page_unlock(m);
 228
 229         return (error);
 230 }
 231
 232 int
 233 uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio)
 234 {
 235         ssize_t resid;
 236         size_t len;
 237         int error;
 238
 239         error = 0;
 240         while ((resid = uio->uio_resid) > 0) {
 241                 if (obj_size <= uio->uio_offset)
 242                         break;
 243                 len = MIN(obj_size - uio->uio_offset, resid);
 244                 if (len == 0)
 245                         break;
 246                 error = uiomove_object_page(obj, len, uio);
 247                 if (error != 0 || resid == uio->uio_resid)
 248                         break;
 249         }
 250         return (error);
 251 }
 252
 253 static int
 254 shm_seek(struct file *fp, off_t offset, int whence, struct thread *td)
 255 {
 256         struct shmfd *shmfd;
 257         off_t foffset;
 258         int error;
 259
 260         shmfd = fp->f_data;
 261         foffset = foffset_lock(fp, 0);
 262         error = 0;
 263         switch (whence) {
 264         case L_INCR:
 265                 if (foffset < 0 ||
 266                     (offset > 0 && foffset > OFF_MAX - offset)) {
 267                         error = EOVERFLOW;
 268                         break;
 269                 }
 270                 offset += foffset;
 271                 break;
 272         case L_XTND:
 273                 if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) {
 274                         error = EOVERFLOW;
 275                         break;
 276                 }
 277                 offset += shmfd->shm_size;
 278                 break;
 279         case L_SET:
 280                 break;
 281         default:
 282                 error = EINVAL;
 283         }
 284         if (error == 0) {
 285                 if (offset < 0 || offset > shmfd->shm_size)
 286                         error = EINVAL;
 287                 else
 288                         td->td_uretoff.tdu_off = offset;
 289         }
 290         foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
 291         return (error);
 292 }
 293
 294 static int
 295 shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
 296     int flags, struct thread *td)
 297 {
 298         struct shmfd *shmfd;
 299         void *rl_cookie;
 300         int error;
 301
 302         shmfd = fp->f_data;
 303 #ifdef MAC
 304         error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd);
 305         if (error)
 306                 return (error);
 307 #endif
 308         foffset_lock_uio(fp, uio, flags);
 309         rl_cookie = rangelock_rlock(&shmfd->shm_rl, uio->uio_offset,
 310             uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
 311         error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
 312         rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
 313         foffset_unlock_uio(fp, uio, flags);
 314         return (error);
 315 }
 316
 317 static int
 318 shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
 319     int flags, struct thread *td)
 320 {
 321         struct shmfd *shmfd;
 322         void *rl_cookie;
 323         int error;
 324
 325         shmfd = fp->f_data;
 326 #ifdef MAC
 327         error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd);
 328         if (error)
 329                 return (error);
 330 #endif
 331         foffset_lock_uio(fp, uio, flags);
 332         if ((flags & FOF_OFFSET) == 0) {
 333                 rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
 334                     &shmfd->shm_mtx);
 335         } else {
 336                 rl_cookie = rangelock_wlock(&shmfd->shm_rl, uio->uio_offset,
 337                     uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
 338         }
 339
 340         error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
 341         rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
 342         foffset_unlock_uio(fp, uio, flags);
 343         return (error);
 344 }
 345
 346 static int
 347 shm_truncate(struct file *fp, off_t length, struct ucred *active_cred,
 348     struct thread *td)
 349 {
 350         struct shmfd *shmfd;
 351 #ifdef MAC
 352         int error;
 353 #endif
 354
 355         shmfd = fp->f_data;
 356 #ifdef MAC
 357         error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd);
 358         if (error)
 359                 return (error);
 360 #endif
 361         return (shm_dotruncate(shmfd, length));
 362 }
 363
 364 int
 365 shm_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
 366     struct thread *td)
 367 {
 368
 369         switch (com) {
 370         case FIONBIO:
 371         case FIOASYNC:
 372                 /*
 373                  * Allow fcntl(fd, F_SETFL, O_NONBLOCK) to work,
 374                  * just like it would on an unlinked regular file
 375                  */
 376                 return (0);
 377         default:
 378                 return (ENOTTY);
 379         }
 380 }
 381
 382 static int
 383 shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred,
 384     struct thread *td)
 385 {
 386         struct shmfd *shmfd;
 387 #ifdef MAC
 388         int error;
 389 #endif
 390
 391         shmfd = fp->f_data;
 392
 393 #ifdef MAC
 394         error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd);
 395         if (error)
 396                 return (error);
 397 #endif
 398
 399         /*
 400          * Attempt to return sanish values for fstat() on a memory file
 401          * descriptor.
 402          */
 403         bzero(sb, sizeof(*sb));
 404         sb->st_blksize = PAGE_SIZE;
 405         sb->st_size = shmfd->shm_size;
 406         sb->st_blocks = howmany(sb->st_size, sb->st_blksize);
 407         mtx_lock(&shm_timestamp_lock);
 408         sb->st_atim = shmfd->shm_atime;
 409         sb->st_ctim = shmfd->shm_ctime;
 410         sb->st_mtim = shmfd->shm_mtime;
 411         sb->st_birthtim = shmfd->shm_birthtime;
 412         sb->st_mode = S_IFREG | shmfd->shm_mode;                /* XXX */
 413         sb->st_uid = shmfd->shm_uid;
 414         sb->st_gid = shmfd->shm_gid;
 415         mtx_unlock(&shm_timestamp_lock);
 416         sb->st_dev = shm_dev_ino;
 417         sb->st_ino = shmfd->shm_ino;
 418         sb->st_nlink = shmfd->shm_object->ref_count;
 419
 420         return (0);
 421 }
 422
 423 static int
 424 shm_close(struct file *fp, struct thread *td)
 425 {
 426         struct shmfd *shmfd;
 427
 428         shmfd = fp->f_data;
 429         fp->f_data = NULL;
 430         shm_drop(shmfd);
 431
 432         return (0);
 433 }
 434
 435 int
 436 shm_dotruncate(struct shmfd *shmfd, off_t length)
 437 {
 438         vm_object_t object;
 439         vm_page_t m;
 440         vm_pindex_t idx, nobjsize;
 441         vm_ooffset_t delta;
 442         int base, rv;
 443
 444         KASSERT(length >= 0, ("shm_dotruncate: length < 0"));
 445         object = shmfd->shm_object;
 446         VM_OBJECT_WLOCK(object);
 447         if (length == shmfd->shm_size) {
 448                 VM_OBJECT_WUNLOCK(object);
 449                 return (0);
 450         }
 451         nobjsize = OFF_TO_IDX(length + PAGE_MASK);
 452
 453         /* Are we shrinking?  If so, trim the end. */
 454         if (length < shmfd->shm_size) {
 455                 /*
 456                  * Disallow any requests to shrink the size if this
 457                  * object is mapped into the kernel.
 458                  */
 459                 if (shmfd->shm_kmappings > 0) {
 460                         VM_OBJECT_WUNLOCK(object);
 461                         return (EBUSY);
 462                 }
 463
 464                 /*
 465                  * Zero the truncated part of the last page.
 466                  */
 467                 base = length & PAGE_MASK;
 468                 if (base != 0) {
 469                         idx = OFF_TO_IDX(length);
 470 retry:
 471                         m = vm_page_lookup(object, idx);
 472                         if (m != NULL) {
 473                                 if (vm_page_sleep_if_busy(m, "shmtrc"))
 474                                         goto retry;
 475                         } else if (vm_pager_has_page(object, idx, NULL, NULL)) {
 476                                 m = vm_page_alloc(object, idx,
 477                                     VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL);
 478                                 if (m == NULL)
 479                                         goto retry;
 480                                 rv = vm_pager_get_pages(object, &m, 1, NULL,
 481                                     NULL);
 482                                 vm_page_lock(m);
 483                                 if (rv == VM_PAGER_OK) {
 484                                         /*
 485                                          * Since the page was not resident,
 486                                          * and therefore not recently
 487                                          * accessed, immediately enqueue it
 488                                          * for asynchronous laundering.  The
 489                                          * current operation is not regarded
 490                                          * as an access.
 491                                          */
 492                                         vm_page_launder(m);
 493                                         vm_page_unlock(m);
 494                                         vm_page_xunbusy(m);
 495                                 } else {
 496                                         vm_page_free(m);
 497                                         vm_page_unlock(m);
 498                                         VM_OBJECT_WUNLOCK(object);
 499                                         return (EIO);
 500                                 }
 501                         }
 502                         if (m != NULL) {
 503                                 pmap_zero_page_area(m, base, PAGE_SIZE - base);
 504                                 KASSERT(m->valid == VM_PAGE_BITS_ALL,
 505                                     ("shm_dotruncate: page %p is invalid", m));
 506                                 vm_page_dirty(m);
 507                                 vm_pager_page_unswapped(m);
 508                         }
 509                 }
 510                 delta = IDX_TO_OFF(object->size - nobjsize);
 511
 512                 /* Toss in memory pages. */
 513                 if (nobjsize < object->size)
 514                         vm_object_page_remove(object, nobjsize, object->size,
 515                             0);
 516
 517                 /* Toss pages from swap. */
 518                 if (object->type == OBJT_SWAP)
 519                         swap_pager_freespace(object, nobjsize, delta);
 520
 521                 /* Free the swap accounted for shm */
 522                 swap_release_by_cred(delta, object->cred);
 523                 object->charge -= delta;
 524         } else {
 525                 /* Try to reserve additional swap space. */
 526                 delta = IDX_TO_OFF(nobjsize - object->size);
 527                 if (!swap_reserve_by_cred(delta, object->cred)) {
 528                         VM_OBJECT_WUNLOCK(object);
 529                         return (ENOMEM);
 530                 }
 531                 object->charge += delta;
 532         }
 533         shmfd->shm_size = length;
 534         shmfd->shm_oldsize = (size_t)length;
 535         mtx_lock(&shm_timestamp_lock);
 536         vfs_timestamp(&shmfd->shm_ctime);
 537         shmfd->shm_mtime = shmfd->shm_ctime;
 538         mtx_unlock(&shm_timestamp_lock);
 539         object->size = nobjsize;
 540         VM_OBJECT_WUNLOCK(object);
 541         return (0);
 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)
 706 {
 707         struct filedesc *fdp;
 708         struct shmfd *shmfd;
 709         struct file *fp;
 710         char *path;
 711         const char *pr_path;
 712         size_t pr_pathlen;
 713         Fnv32_t fnv;
 714         mode_t cmode;
 715         int fd, error;
 716
 717 #ifdef CAPABILITY_MODE
 718         /*
 719          * shm_open(2) is only allowed for anonymous objects.
 720          */
 721         if (IN_CAPABILITY_MODE(td) && (userpath != SHM_ANON))
 722                 return (ECAPMODE);
 723 #endif
 724
 725         AUDIT_ARG_FFLAGS(flags);
 726         AUDIT_ARG_MODE(mode);
 727
 728         if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR)
 729                 return (EINVAL);
 730
 731         if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0)
 732                 return (EINVAL);
 733
 734         fdp = td->td_proc->p_fd;
 735         cmode = (mode & ~fdp->fd_cmask) & ACCESSPERMS;
 736
 737         /*
 738          * shm_open(2) created shm should always have O_CLOEXEC set, as mandated
 739          * by POSIX.  We allow it to be unset here so that an in-kernel
 740          * interface may be written as a thin layer around shm, optionally not
 741          * setting CLOEXEC.  For shm_open(2), O_CLOEXEC is set unconditionally
 742          * in sys_shm_open() to keep this implementation compliant.
 743          */
 744         error = falloc_caps(td, &fp, &fd, flags & O_CLOEXEC, fcaps);
 745         if (error)
 746                 return (error);
 747
 748         /* A SHM_ANON path pointer creates an anonymous object. */
 749         if (userpath == SHM_ANON) {
 750                 /* A read-only anonymous object is pointless. */
 751                 if ((flags & O_ACCMODE) == O_RDONLY) {
 752                         fdclose(td, fp, fd);
 753                         fdrop(fp, td);
 754                         return (EINVAL);
 755                 }
 756                 shmfd = shm_alloc(td->td_ucred, cmode);
 757         } else {
 758                 path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK);
 759                 pr_path = td->td_ucred->cr_prison->pr_path;
 760
 761                 /* Construct a full pathname for jailed callers. */
 762                 pr_pathlen = strcmp(pr_path, "/") == 0 ? 0
 763                     : strlcpy(path, pr_path, MAXPATHLEN);
 764                 error = copyinstr(userpath, path + pr_pathlen,
 765                     MAXPATHLEN - pr_pathlen, NULL);
 766 #ifdef KTRACE
 767                 if (error == 0 && KTRPOINT(curthread, KTR_NAMEI))
 768                         ktrnamei(path);
 769 #endif
 770                 /* Require paths to start with a '/' character. */
 771                 if (error == 0 && path[pr_pathlen] != '/')
 772                         error = EINVAL;
 773                 if (error) {
 774                         fdclose(td, fp, fd);
 775                         fdrop(fp, td);
 776                         free(path, M_SHMFD);
 777                         return (error);
 778                 }
 779
 780                 AUDIT_ARG_UPATH1_CANON(path);
 781                 fnv = fnv_32_str(path, FNV1_32_INIT);
 782                 sx_xlock(&shm_dict_lock);
 783                 shmfd = shm_lookup(path, fnv);
 784                 if (shmfd == NULL) {
 785                         /* Object does not yet exist, create it if requested. */
 786                         if (flags & O_CREAT) {
 787 #ifdef MAC
 788                                 error = mac_posixshm_check_create(td->td_ucred,
 789                                     path);
 790                                 if (error == 0) {
 791 #endif
 792                                         shmfd = shm_alloc(td->td_ucred, cmode);
 793                                         shm_insert(path, fnv, shmfd);
 794 #ifdef MAC
 795                                 }
 796 #endif
 797                         } else {
 798                                 free(path, M_SHMFD);
 799                                 error = ENOENT;
 800                         }
 801                 } else {
 802                         /*
 803                          * Object already exists, obtain a new
 804                          * reference if requested and permitted.
 805                          */
 806                         free(path, M_SHMFD);
 807                         if ((flags & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
 808                                 error = EEXIST;
 809                         else {
 810 #ifdef MAC
 811                                 error = mac_posixshm_check_open(td->td_ucred,
 812                                     shmfd, FFLAGS(flags & O_ACCMODE));
 813                                 if (error == 0)
 814 #endif
 815                                 error = shm_access(shmfd, td->td_ucred,
 816                                     FFLAGS(flags & O_ACCMODE));
 817                         }
 818
 819                         /*
 820                          * Truncate the file back to zero length if
 821                          * O_TRUNC was specified and the object was
 822                          * opened with read/write.
 823                          */
 824                         if (error == 0 &&
 825                             (flags & (O_ACCMODE | O_TRUNC)) ==
 826                             (O_RDWR | O_TRUNC)) {
 827 #ifdef MAC
 828                                 error = mac_posixshm_check_truncate(
 829                                         td->td_ucred, fp->f_cred, shmfd);
 830                                 if (error == 0)
 831 #endif
 832                                         shm_dotruncate(shmfd, 0);
 833                         }
 834                         if (error == 0)
 835                                 shm_hold(shmfd);
 836                 }
 837                 sx_xunlock(&shm_dict_lock);
 838
 839                 if (error) {
 840                         fdclose(td, fp, fd);
 841                         fdrop(fp, td);
 842                         return (error);
 843                 }
 844         }
 845
 846         finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops);
 847
 848         td->td_retval[0] = fd;
 849         fdrop(fp, td);
 850
 851         return (0);
 852 }
 853
 854 /* System calls. */
 855 int
 856 sys_shm_open(struct thread *td, struct shm_open_args *uap)
 857 {
 858
 859         return (kern_shm_open(td, uap->path, uap->flags | O_CLOEXEC, uap->mode,
 860             NULL));
 861 }
 862
 863 int
 864 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap)
 865 {
 866         char *path;
 867         const char *pr_path;
 868         size_t pr_pathlen;
 869         Fnv32_t fnv;
 870         int error;
 871
 872         path = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
 873         pr_path = td->td_ucred->cr_prison->pr_path;
 874         pr_pathlen = strcmp(pr_path, "/") == 0 ? 0
 875             : strlcpy(path, pr_path, MAXPATHLEN);
 876         error = copyinstr(uap->path, path + pr_pathlen, MAXPATHLEN - pr_pathlen,
 877             NULL);
 878         if (error) {
 879                 free(path, M_TEMP);
 880                 return (error);
 881         }
 882 #ifdef KTRACE
 883         if (KTRPOINT(curthread, KTR_NAMEI))
 884                 ktrnamei(path);
 885 #endif
 886         AUDIT_ARG_UPATH1_CANON(path);
 887         fnv = fnv_32_str(path, FNV1_32_INIT);
 888         sx_xlock(&shm_dict_lock);
 889         error = shm_remove(path, fnv, td->td_ucred);
 890         sx_xunlock(&shm_dict_lock);
 891         free(path, M_TEMP);
 892
 893         return (error);
 894 }
 895
 896 int
 897 shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize,
 898     vm_prot_t prot, vm_prot_t cap_maxprot, int flags,
 899     vm_ooffset_t foff, struct thread *td)
 900 {
 901         struct shmfd *shmfd;
 902         vm_prot_t maxprot;
 903         int error;
 904         bool writecnt;
 905
 906         shmfd = fp->f_data;
 907         maxprot = VM_PROT_NONE;
 908
 909         /* FREAD should always be set. */
 910         if ((fp->f_flag & FREAD) != 0)
 911                 maxprot |= VM_PROT_EXECUTE | VM_PROT_READ;
 912         if ((fp->f_flag & FWRITE) != 0)
 913                 maxprot |= VM_PROT_WRITE;
 914
 915         writecnt = (flags & MAP_SHARED) != 0 && (prot & VM_PROT_WRITE) != 0;
 916
 917         /* Don't permit shared writable mappings on read-only descriptors. */
 918         if (writecnt && (maxprot & VM_PROT_WRITE) == 0)
 919                 return (EACCES);
 920         maxprot &= cap_maxprot;
 921
 922         /* See comment in vn_mmap(). */
 923         if (
 924 #ifdef _LP64
 925             objsize > OFF_MAX ||
 926 #endif
 927             foff < 0 || foff > OFF_MAX - objsize)
 928                 return (EINVAL);
 929
 930 #ifdef MAC
 931         error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags);
 932         if (error != 0)
 933                 return (error);
 934 #endif
 935
 936         mtx_lock(&shm_timestamp_lock);
 937         vfs_timestamp(&shmfd->shm_atime);
 938         mtx_unlock(&shm_timestamp_lock);
 939         vm_object_reference(shmfd->shm_object);
 940
 941         if (writecnt)
 942                 vm_pager_update_writecount(shmfd->shm_object, 0, objsize);
 943         error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags,
 944             shmfd->shm_object, foff, writecnt, td);
 945         if (error != 0) {
 946                 if (writecnt)
 947                         vm_pager_release_writecount(shmfd->shm_object, 0,
 948                             objsize);
 949                 vm_object_deallocate(shmfd->shm_object);
 950         }
 951         return (error);
 952 }
 953
 954 static int
 955 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
 956     struct thread *td)
 957 {
 958         struct shmfd *shmfd;
 959         int error;
 960
 961         error = 0;
 962         shmfd = fp->f_data;
 963         mtx_lock(&shm_timestamp_lock);
 964         /*
 965          * SUSv4 says that x bits of permission need not be affected.
 966          * Be consistent with our shm_open there.
 967          */
 968 #ifdef MAC
 969         error = mac_posixshm_check_setmode(active_cred, shmfd, mode);
 970         if (error != 0)
 971                 goto out;
 972 #endif
 973         error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid,
 974             shmfd->shm_gid, VADMIN, active_cred, NULL);
 975         if (error != 0)
 976                 goto out;
 977         shmfd->shm_mode = mode & ACCESSPERMS;
 978 out:
 979         mtx_unlock(&shm_timestamp_lock);
 980         return (error);
 981 }
 982
 983 static int
 984 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
 985     struct thread *td)
 986 {
 987         struct shmfd *shmfd;
 988         int error;
 989
 990         error = 0;
 991         shmfd = fp->f_data;
 992         mtx_lock(&shm_timestamp_lock);
 993 #ifdef MAC
 994         error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid);
 995         if (error != 0)
 996                 goto out;
 997 #endif
 998         if (uid == (uid_t)-1)
 999                 uid = shmfd->shm_uid;
1000         if (gid == (gid_t)-1)
1001                  gid = shmfd->shm_gid;
1002         if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) ||
1003             (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) &&
1004             (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN, 0)))
1005                 goto out;
1006         shmfd->shm_uid = uid;
1007         shmfd->shm_gid = gid;
1008 out:
1009         mtx_unlock(&shm_timestamp_lock);
1010         return (error);
1011 }
1012
1013 /*
1014  * Helper routines to allow the backing object of a shared memory file
1015  * descriptor to be mapped in the kernel.
1016  */
1017 int
1018 shm_map(struct file *fp, size_t size, off_t offset, void **memp)
1019 {
1020         struct shmfd *shmfd;
1021         vm_offset_t kva, ofs;
1022         vm_object_t obj;
1023         int rv;
1024
1025         if (fp->f_type != DTYPE_SHM)
1026                 return (EINVAL);
1027         shmfd = fp->f_data;
1028         obj = shmfd->shm_object;
1029         VM_OBJECT_WLOCK(obj);
1030         /*
1031          * XXXRW: This validation is probably insufficient, and subject to
1032          * sign errors.  It should be fixed.
1033          */
1034         if (offset >= shmfd->shm_size ||
1035             offset + size > round_page(shmfd->shm_size)) {
1036                 VM_OBJECT_WUNLOCK(obj);
1037                 return (EINVAL);
1038         }
1039
1040         shmfd->shm_kmappings++;
1041         vm_object_reference_locked(obj);
1042         VM_OBJECT_WUNLOCK(obj);
1043
1044         /* Map the object into the kernel_map and wire it. */
1045         kva = vm_map_min(kernel_map);
1046         ofs = offset & PAGE_MASK;
1047         offset = trunc_page(offset);
1048         size = round_page(size + ofs);
1049         rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0,
1050             VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE,
1051             VM_PROT_READ | VM_PROT_WRITE, 0);
1052         if (rv == KERN_SUCCESS) {
1053                 rv = vm_map_wire(kernel_map, kva, kva + size,
1054                     VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
1055                 if (rv == KERN_SUCCESS) {
1056                         *memp = (void *)(kva + ofs);
1057                         return (0);
1058                 }
1059                 vm_map_remove(kernel_map, kva, kva + size);
1060         } else
1061                 vm_object_deallocate(obj);
1062
1063         /* On failure, drop our mapping reference. */
1064         VM_OBJECT_WLOCK(obj);
1065         shmfd->shm_kmappings--;
1066         VM_OBJECT_WUNLOCK(obj);
1067
1068         return (vm_mmap_to_errno(rv));
1069 }
1070
1071 /*
1072  * We require the caller to unmap the entire entry.  This allows us to
1073  * safely decrement shm_kmappings when a mapping is removed.
1074  */
1075 int
1076 shm_unmap(struct file *fp, void *mem, size_t size)
1077 {
1078         struct shmfd *shmfd;
1079         vm_map_entry_t entry;
1080         vm_offset_t kva, ofs;
1081         vm_object_t obj;
1082         vm_pindex_t pindex;
1083         vm_prot_t prot;
1084         boolean_t wired;
1085         vm_map_t map;
1086         int rv;
1087
1088         if (fp->f_type != DTYPE_SHM)
1089                 return (EINVAL);
1090         shmfd = fp->f_data;
1091         kva = (vm_offset_t)mem;
1092         ofs = kva & PAGE_MASK;
1093         kva = trunc_page(kva);
1094         size = round_page(size + ofs);
1095         map = kernel_map;
1096         rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry,
1097             &obj, &pindex, &prot, &wired);
1098         if (rv != KERN_SUCCESS)
1099                 return (EINVAL);
1100         if (entry->start != kva || entry->end != kva + size) {
1101                 vm_map_lookup_done(map, entry);
1102                 return (EINVAL);
1103         }
1104         vm_map_lookup_done(map, entry);
1105         if (obj != shmfd->shm_object)
1106                 return (EINVAL);
1107         vm_map_remove(map, kva, kva + size);
1108         VM_OBJECT_WLOCK(obj);
1109         KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped"));
1110         shmfd->shm_kmappings--;
1111         VM_OBJECT_WUNLOCK(obj);
1112         return (0);
1113 }
1114
1115 static int
1116 shm_fill_kinfo_locked(struct shmfd *shmfd, struct kinfo_file *kif, bool list)
1117 {
1118         const char *path, *pr_path;
1119         size_t pr_pathlen;
1120         bool visible;
1121
1122         sx_assert(&shm_dict_lock, SA_LOCKED);
1123         kif->kf_type = KF_TYPE_SHM;
1124         kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode;
1125         kif->kf_un.kf_file.kf_file_size = shmfd->shm_size;
1126         if (shmfd->shm_path != NULL) {
1127                 if (shmfd->shm_path != NULL) {
1128                         path = shmfd->shm_path;
1129                         pr_path = curthread->td_ucred->cr_prison->pr_path;
1130                         if (strcmp(pr_path, "/") != 0) {
1131                                 /* Return the jail-rooted pathname. */
1132                                 pr_pathlen = strlen(pr_path);
1133                                 visible = strncmp(path, pr_path, pr_pathlen)
1134                                     == 0 && path[pr_pathlen] == '/';
1135                                 if (list && !visible)
1136                                         return (EPERM);
1137                                 if (visible)
1138                                         path += pr_pathlen;
1139                         }
1140                         strlcpy(kif->kf_path, path, sizeof(kif->kf_path));
1141                 }
1142         }
1143         return (0);
1144 }
1145
1146 static int
1147 shm_fill_kinfo(struct file *fp, struct kinfo_file *kif,
1148     struct filedesc *fdp __unused)
1149 {
1150         int res;
1151
1152         sx_slock(&shm_dict_lock);
1153         res = shm_fill_kinfo_locked(fp->f_data, kif, false);
1154         sx_sunlock(&shm_dict_lock);
1155         return (res);
1156 }
1157
1158 static int
1159 sysctl_posix_shm_list(SYSCTL_HANDLER_ARGS)
1160 {
1161         struct shm_mapping *shmm;
1162         struct sbuf sb;
1163         struct kinfo_file kif;
1164         u_long i;
1165         ssize_t curlen;
1166         int error, error2;
1167
1168         sbuf_new_for_sysctl(&sb, NULL, sizeof(struct kinfo_file) * 5, req);
1169         sbuf_clear_flags(&sb, SBUF_INCLUDENUL);
1170         curlen = 0;
1171         error = 0;
1172         sx_slock(&shm_dict_lock);
1173         for (i = 0; i < shm_hash + 1; i++) {
1174                 LIST_FOREACH(shmm, &shm_dictionary[i], sm_link) {
1175                         error = shm_fill_kinfo_locked(shmm->sm_shmfd,
1176                             &kif, true);
1177                         if (error == EPERM)
1178                                 continue;
1179                         if (error != 0)
1180                                 break;
1181                         pack_kinfo(&kif);
1182                         if (req->oldptr != NULL &&
1183                             kif.kf_structsize + curlen > req->oldlen)
1184                                 break;
1185                         error = sbuf_bcat(&sb, &kif, kif.kf_structsize) == 0 ?
1186                             0 : ENOMEM;
1187                         if (error != 0)
1188                                 break;
1189                         curlen += kif.kf_structsize;
1190                 }
1191         }
1192         sx_sunlock(&shm_dict_lock);
1193         error2 = sbuf_finish(&sb);
1194         sbuf_delete(&sb);
1195         return (error != 0 ? error : error2);
1196 }
1197
1198 SYSCTL_PROC(_kern_ipc, OID_AUTO, posix_shm_list,
1199     CTLFLAG_RD | CTLFLAG_MPSAFE | CTLTYPE_OPAQUE,
1200     NULL, 0, sysctl_posix_shm_list, "",
1201     "POSIX SHM list");