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/uio.h>
  61 #include <sys/signal.h>
  62 #include <sys/jail.h>
  63 #include <sys/ktrace.h>
  64 #include <sys/lock.h>
  65 #include <sys/malloc.h>
  66 #include <sys/mman.h>
  67 #include <sys/mutex.h>
  68 #include <sys/priv.h>
  69 #include <sys/proc.h>
  70 #include <sys/refcount.h>
  71 #include <sys/resourcevar.h>
  72 #include <sys/rwlock.h>
  73 #include <sys/sbuf.h>
  74 #include <sys/stat.h>
  75 #include <sys/syscallsubr.h>
  76 #include <sys/sysctl.h>
  77 #include <sys/sysproto.h>
  78 #include <sys/systm.h>
  79 #include <sys/sx.h>
  80 #include <sys/time.h>
  81 #include <sys/vnode.h>
  82 #include <sys/unistd.h>
  83 #include <sys/user.h>
  84
  85 #include <security/audit/audit.h>
  86 #include <security/mac/mac_framework.h>
  87
  88 #include <vm/vm.h>
  89 #include <vm/vm_param.h>
  90 #include <vm/pmap.h>
  91 #include <vm/vm_extern.h>
  92 #include <vm/vm_map.h>
  93 #include <vm/vm_kern.h>
  94 #include <vm/vm_object.h>
  95 #include <vm/vm_page.h>
  96 #include <vm/vm_pageout.h>
  97 #include <vm/vm_pager.h>
  98 #include <vm/swap_pager.h>
  99
 100 struct shm_mapping {
 101         char            *sm_path;
 102         Fnv32_t         sm_fnv;
 103         struct shmfd    *sm_shmfd;
 104         LIST_ENTRY(shm_mapping) sm_link;
 105 };
 106
 107 static MALLOC_DEFINE(M_SHMFD, "shmfd", "shared memory file descriptor");
 108 static LIST_HEAD(, shm_mapping) *shm_dictionary;
 109 static struct sx shm_dict_lock;
 110 static struct mtx shm_timestamp_lock;
 111 static u_long shm_hash;
 112 static struct unrhdr64 shm_ino_unr;
 113 static dev_t shm_dev_ino;
 114
 115 #define SHM_HASH(fnv)   (&shm_dictionary[(fnv) & shm_hash])
 116
 117 static void     shm_init(void *arg);
 118 static void     shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd);
 119 static struct shmfd *shm_lookup(char *path, Fnv32_t fnv);
 120 static int      shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred);
 121
 122 static fo_rdwr_t        shm_read;
 123 static fo_rdwr_t        shm_write;
 124 static fo_truncate_t    shm_truncate;
 125 static fo_ioctl_t       shm_ioctl;
 126 static fo_stat_t        shm_stat;
 127 static fo_close_t       shm_close;
 128 static fo_chmod_t       shm_chmod;
 129 static fo_chown_t       shm_chown;
 130 static fo_seek_t        shm_seek;
 131 static fo_fill_kinfo_t  shm_fill_kinfo;
 132 static fo_mmap_t        shm_mmap;
 133
 134 /* File descriptor operations. */
 135 struct fileops shm_ops = {
 136         .fo_read = shm_read,
 137         .fo_write = shm_write,
 138         .fo_truncate = shm_truncate,
 139         .fo_ioctl = shm_ioctl,
 140         .fo_poll = invfo_poll,
 141         .fo_kqfilter = invfo_kqfilter,
 142         .fo_stat = shm_stat,
 143         .fo_close = shm_close,
 144         .fo_chmod = shm_chmod,
 145         .fo_chown = shm_chown,
 146         .fo_sendfile = vn_sendfile,
 147         .fo_seek = shm_seek,
 148         .fo_fill_kinfo = shm_fill_kinfo,
 149         .fo_mmap = shm_mmap,
 150         .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
 151 };
 152
 153 FEATURE(posix_shm, "POSIX shared memory");
 154
 155 static int
 156 uiomove_object_page(vm_object_t obj, size_t len, struct uio *uio)
 157 {
 158         vm_page_t m;
 159         vm_pindex_t idx;
 160         size_t tlen;
 161         int error, offset, rv;
 162
 163         idx = OFF_TO_IDX(uio->uio_offset);
 164         offset = uio->uio_offset & PAGE_MASK;
 165         tlen = MIN(PAGE_SIZE - offset, len);
 166
 167         VM_OBJECT_WLOCK(obj);
 168
 169         /*
 170          * Read I/O without either a corresponding resident page or swap
 171          * page: use zero_region.  This is intended to avoid instantiating
 172          * pages on read from a sparse region.
 173          */
 174         if (uio->uio_rw == UIO_READ && vm_page_lookup(obj, idx) == NULL &&
 175             !vm_pager_has_page(obj, idx, NULL, NULL)) {
 176                 VM_OBJECT_WUNLOCK(obj);
 177                 return (uiomove(__DECONST(void *, zero_region), tlen, uio));
 178         }
 179
 180         /*
 181          * Parallel reads of the page content from disk are prevented
 182          * by exclusive busy.
 183          *
 184          * Although the tmpfs vnode lock is held here, it is
 185          * nonetheless safe to sleep waiting for a free page.  The
 186          * pageout daemon does not need to acquire the tmpfs vnode
 187          * lock to page out tobj's pages because tobj is a OBJT_SWAP
 188          * type object.
 189          */
 190         m = vm_page_grab(obj, idx, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY);
 191         if (m->valid != VM_PAGE_BITS_ALL) {
 192                 vm_page_xbusy(m);
 193                 if (vm_pager_has_page(obj, idx, NULL, NULL)) {
 194                         rv = vm_pager_get_pages(obj, &m, 1, NULL, NULL);
 195                         if (rv != VM_PAGER_OK) {
 196                                 printf(
 197             "uiomove_object: vm_obj %p idx %jd valid %x pager error %d\n",
 198                                     obj, idx, m->valid, rv);
 199                                 vm_page_lock(m);
 200                                 vm_page_free(m);
 201                                 vm_page_unlock(m);
 202                                 VM_OBJECT_WUNLOCK(obj);
 203                                 return (EIO);
 204                         }
 205                 } else
 206                         vm_page_zero_invalid(m, TRUE);
 207                 vm_page_xunbusy(m);
 208         }
 209         vm_page_lock(m);
 210         vm_page_hold(m);
 211         if (vm_page_active(m))
 212                 vm_page_reference(m);
 213         else
 214                 vm_page_activate(m);
 215         vm_page_unlock(m);
 216         VM_OBJECT_WUNLOCK(obj);
 217         error = uiomove_fromphys(&m, offset, tlen, uio);
 218         if (uio->uio_rw == UIO_WRITE && error == 0) {
 219                 VM_OBJECT_WLOCK(obj);
 220                 vm_page_dirty(m);
 221                 vm_pager_page_unswapped(m);
 222                 VM_OBJECT_WUNLOCK(obj);
 223         }
 224         vm_page_lock(m);
 225         vm_page_unhold(m);
 226         vm_page_unlock(m);
 227
 228         return (error);
 229 }
 230
 231 int
 232 uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio)
 233 {
 234         ssize_t resid;
 235         size_t len;
 236         int error;
 237
 238         error = 0;
 239         while ((resid = uio->uio_resid) > 0) {
 240                 if (obj_size <= uio->uio_offset)
 241                         break;
 242                 len = MIN(obj_size - uio->uio_offset, resid);
 243                 if (len == 0)
 244                         break;
 245                 error = uiomove_object_page(obj, len, uio);
 246                 if (error != 0 || resid == uio->uio_resid)
 247                         break;
 248         }
 249         return (error);
 250 }
 251
 252 static int
 253 shm_seek(struct file *fp, off_t offset, int whence, struct thread *td)
 254 {
 255         struct shmfd *shmfd;
 256         off_t foffset;
 257         int error;
 258
 259         shmfd = fp->f_data;
 260         foffset = foffset_lock(fp, 0);
 261         error = 0;
 262         switch (whence) {
 263         case L_INCR:
 264                 if (foffset < 0 ||
 265                     (offset > 0 && foffset > OFF_MAX - offset)) {
 266                         error = EOVERFLOW;
 267                         break;
 268                 }
 269                 offset += foffset;
 270                 break;
 271         case L_XTND:
 272                 if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) {
 273                         error = EOVERFLOW;
 274                         break;
 275                 }
 276                 offset += shmfd->shm_size;
 277                 break;
 278         case L_SET:
 279                 break;
 280         default:
 281                 error = EINVAL;
 282         }
 283         if (error == 0) {
 284                 if (offset < 0 || offset > shmfd->shm_size)
 285                         error = EINVAL;
 286                 else
 287                         td->td_uretoff.tdu_off = offset;
 288         }
 289         foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
 290         return (error);
 291 }
 292
 293 static int
 294 shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
 295     int flags, struct thread *td)
 296 {
 297         struct shmfd *shmfd;
 298         void *rl_cookie;
 299         int error;
 300
 301         shmfd = fp->f_data;
 302 #ifdef MAC
 303         error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd);
 304         if (error)
 305                 return (error);
 306 #endif
 307         foffset_lock_uio(fp, uio, flags);
 308         rl_cookie = rangelock_rlock(&shmfd->shm_rl, uio->uio_offset,
 309             uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
 310         error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
 311         rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
 312         foffset_unlock_uio(fp, uio, flags);
 313         return (error);
 314 }
 315
 316 static int
 317 shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
 318     int flags, struct thread *td)
 319 {
 320         struct shmfd *shmfd;
 321         void *rl_cookie;
 322         int error;
 323
 324         shmfd = fp->f_data;
 325 #ifdef MAC
 326         error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd);
 327         if (error)
 328                 return (error);
 329 #endif
 330         foffset_lock_uio(fp, uio, flags);
 331         if ((flags & FOF_OFFSET) == 0) {
 332                 rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
 333                     &shmfd->shm_mtx);
 334         } else {
 335                 rl_cookie = rangelock_wlock(&shmfd->shm_rl, uio->uio_offset,
 336                     uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
 337         }
 338
 339         error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
 340         rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
 341         foffset_unlock_uio(fp, uio, flags);
 342         return (error);
 343 }
 344
 345 static int
 346 shm_truncate(struct file *fp, off_t length, struct ucred *active_cred,
 347     struct thread *td)
 348 {
 349         struct shmfd *shmfd;
 350 #ifdef MAC
 351         int error;
 352 #endif
 353
 354         shmfd = fp->f_data;
 355 #ifdef MAC
 356         error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd);
 357         if (error)
 358                 return (error);
 359 #endif
 360         return (shm_dotruncate(shmfd, length));
 361 }
 362
 363 int
 364 shm_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
 365     struct thread *td)
 366 {
 367
 368         switch (com) {
 369         case FIONBIO:
 370         case FIOASYNC:
 371                 /*
 372                  * Allow fcntl(fd, F_SETFL, O_NONBLOCK) to work,
 373                  * just like it would on an unlinked regular file
 374                  */
 375                 return (0);
 376         default:
 377                 return (ENOTTY);
 378         }
 379 }
 380
 381 static int
 382 shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred,
 383     struct thread *td)
 384 {
 385         struct shmfd *shmfd;
 386 #ifdef MAC
 387         int error;
 388 #endif
 389
 390         shmfd = fp->f_data;
 391
 392 #ifdef MAC
 393         error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd);
 394         if (error)
 395                 return (error);
 396 #endif
 397
 398         /*
 399          * Attempt to return sanish values for fstat() on a memory file
 400          * descriptor.
 401          */
 402         bzero(sb, sizeof(*sb));
 403         sb->st_blksize = PAGE_SIZE;
 404         sb->st_size = shmfd->shm_size;
 405         sb->st_blocks = howmany(sb->st_size, sb->st_blksize);
 406         mtx_lock(&shm_timestamp_lock);
 407         sb->st_atim = shmfd->shm_atime;
 408         sb->st_ctim = shmfd->shm_ctime;
 409         sb->st_mtim = shmfd->shm_mtime;
 410         sb->st_birthtim = shmfd->shm_birthtime;
 411         sb->st_mode = S_IFREG | shmfd->shm_mode;                /* XXX */
 412         sb->st_uid = shmfd->shm_uid;
 413         sb->st_gid = shmfd->shm_gid;
 414         mtx_unlock(&shm_timestamp_lock);
 415         sb->st_dev = shm_dev_ino;
 416         sb->st_ino = shmfd->shm_ino;
 417         sb->st_nlink = shmfd->shm_object->ref_count;
 418
 419         return (0);
 420 }
 421
 422 static int
 423 shm_close(struct file *fp, struct thread *td)
 424 {
 425         struct shmfd *shmfd;
 426
 427         shmfd = fp->f_data;
 428         fp->f_data = NULL;
 429         shm_drop(shmfd);
 430
 431         return (0);
 432 }
 433
 434 int
 435 shm_dotruncate(struct shmfd *shmfd, off_t length)
 436 {
 437         vm_object_t object;
 438         vm_page_t m;
 439         vm_pindex_t idx, nobjsize;
 440         vm_ooffset_t delta;
 441         int base, rv;
 442
 443         KASSERT(length >= 0, ("shm_dotruncate: length < 0"));
 444         object = shmfd->shm_object;
 445         VM_OBJECT_WLOCK(object);
 446         if (length == shmfd->shm_size) {
 447                 VM_OBJECT_WUNLOCK(object);
 448                 return (0);
 449         }
 450         nobjsize = OFF_TO_IDX(length + PAGE_MASK);
 451
 452         /* Are we shrinking?  If so, trim the end. */
 453         if (length < shmfd->shm_size) {
 454                 /*
 455                  * Disallow any requests to shrink the size if this
 456                  * object is mapped into the kernel.
 457                  */
 458                 if (shmfd->shm_kmappings > 0) {
 459                         VM_OBJECT_WUNLOCK(object);
 460                         return (EBUSY);
 461                 }
 462
 463                 /*
 464                  * Zero the truncated part of the last page.
 465                  */
 466                 base = length & PAGE_MASK;
 467                 if (base != 0) {
 468                         idx = OFF_TO_IDX(length);
 469 retry:
 470                         m = vm_page_lookup(object, idx);
 471                         if (m != NULL) {
 472                                 if (vm_page_sleep_if_busy(m, "shmtrc"))
 473                                         goto retry;
 474                         } else if (vm_pager_has_page(object, idx, NULL, NULL)) {
 475                                 m = vm_page_alloc(object, idx,
 476                                     VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL);
 477                                 if (m == NULL)
 478                                         goto retry;
 479                                 rv = vm_pager_get_pages(object, &m, 1, NULL,
 480                                     NULL);
 481                                 vm_page_lock(m);
 482                                 if (rv == VM_PAGER_OK) {
 483                                         /*
 484                                          * Since the page was not resident,
 485                                          * and therefore not recently
 486                                          * accessed, immediately enqueue it
 487                                          * for asynchronous laundering.  The
 488                                          * current operation is not regarded
 489                                          * as an access.
 490                                          */
 491                                         vm_page_launder(m);
 492                                         vm_page_unlock(m);
 493                                         vm_page_xunbusy(m);
 494                                 } else {
 495                                         vm_page_free(m);
 496                                         vm_page_unlock(m);
 497                                         VM_OBJECT_WUNLOCK(object);
 498                                         return (EIO);
 499                                 }
 500                         }
 501                         if (m != NULL) {
 502                                 pmap_zero_page_area(m, base, PAGE_SIZE - base);
 503                                 KASSERT(m->valid == VM_PAGE_BITS_ALL,
 504                                     ("shm_dotruncate: page %p is invalid", m));
 505                                 vm_page_dirty(m);
 506                                 vm_pager_page_unswapped(m);
 507                         }
 508                 }
 509                 delta = IDX_TO_OFF(object->size - nobjsize);
 510
 511                 /* Toss in memory pages. */
 512                 if (nobjsize < object->size)
 513                         vm_object_page_remove(object, nobjsize, object->size,
 514                             0);
 515
 516                 /* Toss pages from swap. */
 517                 if (object->type == OBJT_SWAP)
 518                         swap_pager_freespace(object, nobjsize, delta);
 519
 520                 /* Free the swap accounted for shm */
 521                 swap_release_by_cred(delta, object->cred);
 522                 object->charge -= delta;
 523         } else {
 524                 /* Try to reserve additional swap space. */
 525                 delta = IDX_TO_OFF(nobjsize - object->size);
 526                 if (!swap_reserve_by_cred(delta, object->cred)) {
 527                         VM_OBJECT_WUNLOCK(object);
 528                         return (ENOMEM);
 529                 }
 530                 object->charge += delta;
 531         }
 532         shmfd->shm_size = length;
 533         mtx_lock(&shm_timestamp_lock);
 534         vfs_timestamp(&shmfd->shm_ctime);
 535         shmfd->shm_mtime = shmfd->shm_ctime;
 536         mtx_unlock(&shm_timestamp_lock);
 537         object->size = nobjsize;
 538         VM_OBJECT_WUNLOCK(object);
 539         return (0);
 540 }
 541
 542 /*
 543  * shmfd object management including creation and reference counting
 544  * routines.
 545  */
 546 struct shmfd *
 547 shm_alloc(struct ucred *ucred, mode_t mode)
 548 {
 549         struct shmfd *shmfd;
 550
 551         shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO);
 552         shmfd->shm_size = 0;
 553         shmfd->shm_uid = ucred->cr_uid;
 554         shmfd->shm_gid = ucred->cr_gid;
 555         shmfd->shm_mode = mode;
 556         shmfd->shm_object = vm_pager_allocate(OBJT_DEFAULT, NULL,
 557             shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred);
 558         KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate"));
 559         shmfd->shm_object->pg_color = 0;
 560         VM_OBJECT_WLOCK(shmfd->shm_object);
 561         vm_object_clear_flag(shmfd->shm_object, OBJ_ONEMAPPING);
 562         vm_object_set_flag(shmfd->shm_object, OBJ_COLORED | OBJ_NOSPLIT);
 563         VM_OBJECT_WUNLOCK(shmfd->shm_object);
 564         vfs_timestamp(&shmfd->shm_birthtime);
 565         shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime =
 566             shmfd->shm_birthtime;
 567         shmfd->shm_ino = alloc_unr64(&shm_ino_unr);
 568         refcount_init(&shmfd->shm_refs, 1);
 569         mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF);
 570         rangelock_init(&shmfd->shm_rl);
 571 #ifdef MAC
 572         mac_posixshm_init(shmfd);
 573         mac_posixshm_create(ucred, shmfd);
 574 #endif
 575
 576         return (shmfd);
 577 }
 578
 579 struct shmfd *
 580 shm_hold(struct shmfd *shmfd)
 581 {
 582
 583         refcount_acquire(&shmfd->shm_refs);
 584         return (shmfd);
 585 }
 586
 587 void
 588 shm_drop(struct shmfd *shmfd)
 589 {
 590
 591         if (refcount_release(&shmfd->shm_refs)) {
 592 #ifdef MAC
 593                 mac_posixshm_destroy(shmfd);
 594 #endif
 595                 rangelock_destroy(&shmfd->shm_rl);
 596                 mtx_destroy(&shmfd->shm_mtx);
 597                 vm_object_deallocate(shmfd->shm_object);
 598                 free(shmfd, M_SHMFD);
 599         }
 600 }
 601
 602 /*
 603  * Determine if the credentials have sufficient permissions for a
 604  * specified combination of FREAD and FWRITE.
 605  */
 606 int
 607 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags)
 608 {
 609         accmode_t accmode;
 610         int error;
 611
 612         accmode = 0;
 613         if (flags & FREAD)
 614                 accmode |= VREAD;
 615         if (flags & FWRITE)
 616                 accmode |= VWRITE;
 617         mtx_lock(&shm_timestamp_lock);
 618         error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
 619             accmode, ucred, NULL);
 620         mtx_unlock(&shm_timestamp_lock);
 621         return (error);
 622 }
 623
 624 /*
 625  * Dictionary management.  We maintain an in-kernel dictionary to map
 626  * paths to shmfd objects.  We use the FNV hash on the path to store
 627  * the mappings in a hash table.
 628  */
 629 static void
 630 shm_init(void *arg)
 631 {
 632
 633         mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF);
 634         sx_init(&shm_dict_lock, "shm dictionary");
 635         shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash);
 636         new_unrhdr64(&shm_ino_unr, 1);
 637         shm_dev_ino = devfs_alloc_cdp_inode();
 638         KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized"));
 639 }
 640 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL);
 641
 642 static struct shmfd *
 643 shm_lookup(char *path, Fnv32_t fnv)
 644 {
 645         struct shm_mapping *map;
 646
 647         LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
 648                 if (map->sm_fnv != fnv)
 649                         continue;
 650                 if (strcmp(map->sm_path, path) == 0)
 651                         return (map->sm_shmfd);
 652         }
 653
 654         return (NULL);
 655 }
 656
 657 static void
 658 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd)
 659 {
 660         struct shm_mapping *map;
 661
 662         map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK);
 663         map->sm_path = path;
 664         map->sm_fnv = fnv;
 665         map->sm_shmfd = shm_hold(shmfd);
 666         shmfd->shm_path = path;
 667         LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link);
 668 }
 669
 670 static int
 671 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred)
 672 {
 673         struct shm_mapping *map;
 674         int error;
 675
 676         LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
 677                 if (map->sm_fnv != fnv)
 678                         continue;
 679                 if (strcmp(map->sm_path, path) == 0) {
 680 #ifdef MAC
 681                         error = mac_posixshm_check_unlink(ucred, map->sm_shmfd);
 682                         if (error)
 683                                 return (error);
 684 #endif
 685                         error = shm_access(map->sm_shmfd, ucred,
 686                             FREAD | FWRITE);
 687                         if (error)
 688                                 return (error);
 689                         map->sm_shmfd->shm_path = NULL;
 690                         LIST_REMOVE(map, sm_link);
 691                         shm_drop(map->sm_shmfd);
 692                         free(map->sm_path, M_SHMFD);
 693                         free(map, M_SHMFD);
 694                         return (0);
 695                 }
 696         }
 697
 698         return (ENOENT);
 699 }
 700
 701 int
 702 kern_shm_open(struct thread *td, const char *userpath, int flags, mode_t mode,
 703     struct filecaps *fcaps)
 704 {
 705         struct filedesc *fdp;
 706         struct shmfd *shmfd;
 707         struct file *fp;
 708         char *path;
 709         const char *pr_path;
 710         size_t pr_pathlen;
 711         Fnv32_t fnv;
 712         mode_t cmode;
 713         int fd, error;
 714
 715 #ifdef CAPABILITY_MODE
 716         /*
 717          * shm_open(2) is only allowed for anonymous objects.
 718          */
 719         if (IN_CAPABILITY_MODE(td) && (userpath != SHM_ANON))
 720                 return (ECAPMODE);
 721 #endif
 722
 723         AUDIT_ARG_FFLAGS(flags);
 724         AUDIT_ARG_MODE(mode);
 725
 726         if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR)
 727                 return (EINVAL);
 728
 729         if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0)
 730                 return (EINVAL);
 731
 732         fdp = td->td_proc->p_fd;
 733         cmode = (mode & ~fdp->fd_cmask) & ACCESSPERMS;
 734
 735         error = falloc_caps(td, &fp, &fd, O_CLOEXEC, fcaps);
 736         if (error)
 737                 return (error);
 738
 739         /* A SHM_ANON path pointer creates an anonymous object. */
 740         if (userpath == SHM_ANON) {
 741                 /* A read-only anonymous object is pointless. */
 742                 if ((flags & O_ACCMODE) == O_RDONLY) {
 743                         fdclose(td, fp, fd);
 744                         fdrop(fp, td);
 745                         return (EINVAL);
 746                 }
 747                 shmfd = shm_alloc(td->td_ucred, cmode);
 748         } else {
 749                 path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK);
 750                 pr_path = td->td_ucred->cr_prison->pr_path;
 751
 752                 /* Construct a full pathname for jailed callers. */
 753                 pr_pathlen = strcmp(pr_path, "/") == 0 ? 0
 754                     : strlcpy(path, pr_path, MAXPATHLEN);
 755                 error = copyinstr(userpath, path + pr_pathlen,
 756                     MAXPATHLEN - pr_pathlen, NULL);
 757 #ifdef KTRACE
 758                 if (error == 0 && KTRPOINT(curthread, KTR_NAMEI))
 759                         ktrnamei(path);
 760 #endif
 761                 /* Require paths to start with a '/' character. */
 762                 if (error == 0 && path[pr_pathlen] != '/')
 763                         error = EINVAL;
 764                 if (error) {
 765                         fdclose(td, fp, fd);
 766                         fdrop(fp, td);
 767                         free(path, M_SHMFD);
 768                         return (error);
 769                 }
 770
 771                 AUDIT_ARG_UPATH1_CANON(path);
 772                 fnv = fnv_32_str(path, FNV1_32_INIT);
 773                 sx_xlock(&shm_dict_lock);
 774                 shmfd = shm_lookup(path, fnv);
 775                 if (shmfd == NULL) {
 776                         /* Object does not yet exist, create it if requested. */
 777                         if (flags & O_CREAT) {
 778 #ifdef MAC
 779                                 error = mac_posixshm_check_create(td->td_ucred,
 780                                     path);
 781                                 if (error == 0) {
 782 #endif
 783                                         shmfd = shm_alloc(td->td_ucred, cmode);
 784                                         shm_insert(path, fnv, shmfd);
 785 #ifdef MAC
 786                                 }
 787 #endif
 788                         } else {
 789                                 free(path, M_SHMFD);
 790                                 error = ENOENT;
 791                         }
 792                 } else {
 793                         /*
 794                          * Object already exists, obtain a new
 795                          * reference if requested and permitted.
 796                          */
 797                         free(path, M_SHMFD);
 798                         if ((flags & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
 799                                 error = EEXIST;
 800                         else {
 801 #ifdef MAC
 802                                 error = mac_posixshm_check_open(td->td_ucred,
 803                                     shmfd, FFLAGS(flags & O_ACCMODE));
 804                                 if (error == 0)
 805 #endif
 806                                 error = shm_access(shmfd, td->td_ucred,
 807                                     FFLAGS(flags & O_ACCMODE));
 808                         }
 809
 810                         /*
 811                          * Truncate the file back to zero length if
 812                          * O_TRUNC was specified and the object was
 813                          * opened with read/write.
 814                          */
 815                         if (error == 0 &&
 816                             (flags & (O_ACCMODE | O_TRUNC)) ==
 817                             (O_RDWR | O_TRUNC)) {
 818 #ifdef MAC
 819                                 error = mac_posixshm_check_truncate(
 820                                         td->td_ucred, fp->f_cred, shmfd);
 821                                 if (error == 0)
 822 #endif
 823                                         shm_dotruncate(shmfd, 0);
 824                         }
 825                         if (error == 0)
 826                                 shm_hold(shmfd);
 827                 }
 828                 sx_xunlock(&shm_dict_lock);
 829
 830                 if (error) {
 831                         fdclose(td, fp, fd);
 832                         fdrop(fp, td);
 833                         return (error);
 834                 }
 835         }
 836
 837         finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops);
 838
 839         td->td_retval[0] = fd;
 840         fdrop(fp, td);
 841
 842         return (0);
 843 }
 844
 845 /* System calls. */
 846 int
 847 sys_shm_open(struct thread *td, struct shm_open_args *uap)
 848 {
 849
 850         return (kern_shm_open(td, uap->path, uap->flags, uap->mode, NULL));
 851 }
 852
 853 int
 854 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap)
 855 {
 856         char *path;
 857         const char *pr_path;
 858         size_t pr_pathlen;
 859         Fnv32_t fnv;
 860         int error;
 861
 862         path = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
 863         pr_path = td->td_ucred->cr_prison->pr_path;
 864         pr_pathlen = strcmp(pr_path, "/") == 0 ? 0
 865             : strlcpy(path, pr_path, MAXPATHLEN);
 866         error = copyinstr(uap->path, path + pr_pathlen, MAXPATHLEN - pr_pathlen,
 867             NULL);
 868         if (error) {
 869                 free(path, M_TEMP);
 870                 return (error);
 871         }
 872 #ifdef KTRACE
 873         if (KTRPOINT(curthread, KTR_NAMEI))
 874                 ktrnamei(path);
 875 #endif
 876         AUDIT_ARG_UPATH1_CANON(path);
 877         fnv = fnv_32_str(path, FNV1_32_INIT);
 878         sx_xlock(&shm_dict_lock);
 879         error = shm_remove(path, fnv, td->td_ucred);
 880         sx_xunlock(&shm_dict_lock);
 881         free(path, M_TEMP);
 882
 883         return (error);
 884 }
 885
 886 int
 887 shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize,
 888     vm_prot_t prot, vm_prot_t cap_maxprot, int flags,
 889     vm_ooffset_t foff, struct thread *td)
 890 {
 891         struct shmfd *shmfd;
 892         vm_prot_t maxprot;
 893         int error;
 894
 895         shmfd = fp->f_data;
 896         maxprot = VM_PROT_NONE;
 897
 898         /* FREAD should always be set. */
 899         if ((fp->f_flag & FREAD) != 0)
 900                 maxprot |= VM_PROT_EXECUTE | VM_PROT_READ;
 901         if ((fp->f_flag & FWRITE) != 0)
 902                 maxprot |= VM_PROT_WRITE;
 903
 904         /* Don't permit shared writable mappings on read-only descriptors. */
 905         if ((flags & MAP_SHARED) != 0 &&
 906             (maxprot & VM_PROT_WRITE) == 0 &&
 907             (prot & VM_PROT_WRITE) != 0)
 908                 return (EACCES);
 909         maxprot &= cap_maxprot;
 910
 911         /* See comment in vn_mmap(). */
 912         if (
 913 #ifdef _LP64
 914             objsize > OFF_MAX ||
 915 #endif
 916             foff < 0 || foff > OFF_MAX - objsize)
 917                 return (EINVAL);
 918
 919 #ifdef MAC
 920         error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags);
 921         if (error != 0)
 922                 return (error);
 923 #endif
 924
 925         mtx_lock(&shm_timestamp_lock);
 926         vfs_timestamp(&shmfd->shm_atime);
 927         mtx_unlock(&shm_timestamp_lock);
 928         vm_object_reference(shmfd->shm_object);
 929
 930         error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags,
 931             shmfd->shm_object, foff, FALSE, td);
 932         if (error != 0)
 933                 vm_object_deallocate(shmfd->shm_object);
 934         return (error);
 935 }
 936
 937 static int
 938 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
 939     struct thread *td)
 940 {
 941         struct shmfd *shmfd;
 942         int error;
 943
 944         error = 0;
 945         shmfd = fp->f_data;
 946         mtx_lock(&shm_timestamp_lock);
 947         /*
 948          * SUSv4 says that x bits of permission need not be affected.
 949          * Be consistent with our shm_open there.
 950          */
 951 #ifdef MAC
 952         error = mac_posixshm_check_setmode(active_cred, shmfd, mode);
 953         if (error != 0)
 954                 goto out;
 955 #endif
 956         error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid,
 957             shmfd->shm_gid, VADMIN, active_cred, NULL);
 958         if (error != 0)
 959                 goto out;
 960         shmfd->shm_mode = mode & ACCESSPERMS;
 961 out:
 962         mtx_unlock(&shm_timestamp_lock);
 963         return (error);
 964 }
 965
 966 static int
 967 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
 968     struct thread *td)
 969 {
 970         struct shmfd *shmfd;
 971         int error;
 972
 973         error = 0;
 974         shmfd = fp->f_data;
 975         mtx_lock(&shm_timestamp_lock);
 976 #ifdef MAC
 977         error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid);
 978         if (error != 0)
 979                 goto out;
 980 #endif
 981         if (uid == (uid_t)-1)
 982                 uid = shmfd->shm_uid;
 983         if (gid == (gid_t)-1)
 984                  gid = shmfd->shm_gid;
 985         if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) ||
 986             (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) &&
 987             (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN)))
 988                 goto out;
 989         shmfd->shm_uid = uid;
 990         shmfd->shm_gid = gid;
 991 out:
 992         mtx_unlock(&shm_timestamp_lock);
 993         return (error);
 994 }
 995
 996 /*
 997  * Helper routines to allow the backing object of a shared memory file
 998  * descriptor to be mapped in the kernel.
 999  */
1000 int
1001 shm_map(struct file *fp, size_t size, off_t offset, void **memp)
1002 {
1003         struct shmfd *shmfd;
1004         vm_offset_t kva, ofs;
1005         vm_object_t obj;
1006         int rv;
1007
1008         if (fp->f_type != DTYPE_SHM)
1009                 return (EINVAL);
1010         shmfd = fp->f_data;
1011         obj = shmfd->shm_object;
1012         VM_OBJECT_WLOCK(obj);
1013         /*
1014          * XXXRW: This validation is probably insufficient, and subject to
1015          * sign errors.  It should be fixed.
1016          */
1017         if (offset >= shmfd->shm_size ||
1018             offset + size > round_page(shmfd->shm_size)) {
1019                 VM_OBJECT_WUNLOCK(obj);
1020                 return (EINVAL);
1021         }
1022
1023         shmfd->shm_kmappings++;
1024         vm_object_reference_locked(obj);
1025         VM_OBJECT_WUNLOCK(obj);
1026
1027         /* Map the object into the kernel_map and wire it. */
1028         kva = vm_map_min(kernel_map);
1029         ofs = offset & PAGE_MASK;
1030         offset = trunc_page(offset);
1031         size = round_page(size + ofs);
1032         rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0,
1033             VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE,
1034             VM_PROT_READ | VM_PROT_WRITE, 0);
1035         if (rv == KERN_SUCCESS) {
1036                 rv = vm_map_wire(kernel_map, kva, kva + size,
1037                     VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
1038                 if (rv == KERN_SUCCESS) {
1039                         *memp = (void *)(kva + ofs);
1040                         return (0);
1041                 }
1042                 vm_map_remove(kernel_map, kva, kva + size);
1043         } else
1044                 vm_object_deallocate(obj);
1045
1046         /* On failure, drop our mapping reference. */
1047         VM_OBJECT_WLOCK(obj);
1048         shmfd->shm_kmappings--;
1049         VM_OBJECT_WUNLOCK(obj);
1050
1051         return (vm_mmap_to_errno(rv));
1052 }
1053
1054 /*
1055  * We require the caller to unmap the entire entry.  This allows us to
1056  * safely decrement shm_kmappings when a mapping is removed.
1057  */
1058 int
1059 shm_unmap(struct file *fp, void *mem, size_t size)
1060 {
1061         struct shmfd *shmfd;
1062         vm_map_entry_t entry;
1063         vm_offset_t kva, ofs;
1064         vm_object_t obj;
1065         vm_pindex_t pindex;
1066         vm_prot_t prot;
1067         boolean_t wired;
1068         vm_map_t map;
1069         int rv;
1070
1071         if (fp->f_type != DTYPE_SHM)
1072                 return (EINVAL);
1073         shmfd = fp->f_data;
1074         kva = (vm_offset_t)mem;
1075         ofs = kva & PAGE_MASK;
1076         kva = trunc_page(kva);
1077         size = round_page(size + ofs);
1078         map = kernel_map;
1079         rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry,
1080             &obj, &pindex, &prot, &wired);
1081         if (rv != KERN_SUCCESS)
1082                 return (EINVAL);
1083         if (entry->start != kva || entry->end != kva + size) {
1084                 vm_map_lookup_done(map, entry);
1085                 return (EINVAL);
1086         }
1087         vm_map_lookup_done(map, entry);
1088         if (obj != shmfd->shm_object)
1089                 return (EINVAL);
1090         vm_map_remove(map, kva, kva + size);
1091         VM_OBJECT_WLOCK(obj);
1092         KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped"));
1093         shmfd->shm_kmappings--;
1094         VM_OBJECT_WUNLOCK(obj);
1095         return (0);
1096 }
1097
1098 static int
1099 shm_fill_kinfo_locked(struct shmfd *shmfd, struct kinfo_file *kif, bool list)
1100 {
1101         const char *path, *pr_path;
1102         size_t pr_pathlen;
1103         bool visible;
1104
1105         sx_assert(&shm_dict_lock, SA_LOCKED);
1106         kif->kf_type = KF_TYPE_SHM;
1107         kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode;
1108         kif->kf_un.kf_file.kf_file_size = shmfd->shm_size;
1109         if (shmfd->shm_path != NULL) {
1110                 if (shmfd->shm_path != NULL) {
1111                         path = shmfd->shm_path;
1112                         pr_path = curthread->td_ucred->cr_prison->pr_path;
1113                         if (strcmp(pr_path, "/") != 0) {
1114                                 /* Return the jail-rooted pathname. */
1115                                 pr_pathlen = strlen(pr_path);
1116                                 visible = strncmp(path, pr_path, pr_pathlen)
1117                                     == 0 && path[pr_pathlen] == '/';
1118                                 if (list && !visible)
1119                                         return (EPERM);
1120                                 if (visible)
1121                                         path += pr_pathlen;
1122                         }
1123                         strlcpy(kif->kf_path, path, sizeof(kif->kf_path));
1124                 }
1125         }
1126         return (0);
1127 }
1128
1129 static int
1130 shm_fill_kinfo(struct file *fp, struct kinfo_file *kif,
1131     struct filedesc *fdp __unused)
1132 {
1133         int res;
1134
1135         sx_slock(&shm_dict_lock);
1136         res = shm_fill_kinfo_locked(fp->f_data, kif, false);
1137         sx_sunlock(&shm_dict_lock);
1138         return (res);
1139 }
1140
1141 static int
1142 sysctl_posix_shm_list(SYSCTL_HANDLER_ARGS)
1143 {
1144         struct shm_mapping *shmm;
1145         struct sbuf sb;
1146         struct kinfo_file kif;
1147         u_long i;
1148         ssize_t curlen;
1149         int error, error2;
1150
1151         sbuf_new_for_sysctl(&sb, NULL, sizeof(struct kinfo_file) * 5, req);
1152         sbuf_clear_flags(&sb, SBUF_INCLUDENUL);
1153         curlen = 0;
1154         error = 0;
1155         sx_slock(&shm_dict_lock);
1156         for (i = 0; i < shm_hash + 1; i++) {
1157                 LIST_FOREACH(shmm, &shm_dictionary[i], sm_link) {
1158                         error = shm_fill_kinfo_locked(shmm->sm_shmfd,
1159                             &kif, true);
1160                         if (error == EPERM)
1161                                 continue;
1162                         if (error != 0)
1163                                 break;
1164                         pack_kinfo(&kif);
1165                         if (req->oldptr != NULL &&
1166                             kif.kf_structsize + curlen > req->oldlen)
1167                                 break;
1168                         error = sbuf_bcat(&sb, &kif, kif.kf_structsize) == 0 ?
1169                             0 : ENOMEM;
1170                         if (error != 0)
1171                                 break;
1172                         curlen += kif.kf_structsize;
1173                 }
1174         }
1175         sx_sunlock(&shm_dict_lock);
1176         error2 = sbuf_finish(&sb);
1177         sbuf_delete(&sb);
1178         return (error != 0 ? error : error2);
1179 }
1180
1181 SYSCTL_PROC(_kern_ipc, OID_AUTO, posix_shm_list,
1182     CTLFLAG_RD | CTLFLAG_MPSAFE | CTLTYPE_OPAQUE,
1183     NULL, 0, sysctl_posix_shm_list, "",
1184     "POSIX SHM list");