sys/kern/uipc_shm.c

   1 /*-
   2  * Copyright (c) 2006, 2011 Robert N. M. Watson
   3  * All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice, this list of conditions and the following disclaimer.
  10  * 2. Redistributions in binary form must reproduce the above copyright
  11  *    notice, this list of conditions and the following disclaimer in the
  12  *    documentation and/or other materials provided with the distribution.
  13  *
  14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  24  * SUCH DAMAGE.
  25  */
  26
  27 /*
  28  * Support for shared swap-backed anonymous memory objects via
  29  * shm_open(2) and shm_unlink(2).  While most of the implementation is
  30  * here, vm_mmap.c contains mapping logic changes.
  31  *
  32  * TODO:
  33  *
  34  * (1) Need to export data to a userland tool via a sysctl.  Should ipcs(1)
  35  *     and ipcrm(1) be expanded or should new tools to manage both POSIX
  36  *     kernel semaphores and POSIX shared memory be written?
  37  *
  38  * (2) Add support for this file type to fstat(1).
  39  *
  40  * (3) Resource limits?  Does this need its own resource limits or are the
  41  *     existing limits in mmap(2) sufficient?
  42  */
  43
  44 #include <sys/cdefs.h>
  45 __FBSDID("$FreeBSD$");
  46
  47 #include "opt_capsicum.h"
  48 #include "opt_ktrace.h"
  49
  50 #include <sys/param.h>
  51 #include <sys/capability.h>
  52 #include <sys/conf.h>
  53 #include <sys/fcntl.h>
  54 #include <sys/file.h>
  55 #include <sys/filedesc.h>
  56 #include <sys/fnv_hash.h>
  57 #include <sys/kernel.h>
  58 #include <sys/uio.h>
  59 #include <sys/signal.h>
  60 #include <sys/ktrace.h>
  61 #include <sys/lock.h>
  62 #include <sys/malloc.h>
  63 #include <sys/mman.h>
  64 #include <sys/mutex.h>
  65 #include <sys/priv.h>
  66 #include <sys/proc.h>
  67 #include <sys/refcount.h>
  68 #include <sys/resourcevar.h>
  69 #include <sys/rwlock.h>
  70 #include <sys/stat.h>
  71 #include <sys/sysctl.h>
  72 #include <sys/sysproto.h>
  73 #include <sys/systm.h>
  74 #include <sys/sx.h>
  75 #include <sys/time.h>
  76 #include <sys/vnode.h>
  77 #include <sys/unistd.h>
  78
  79 #include <security/mac/mac_framework.h>
  80
  81 #include <vm/vm.h>
  82 #include <vm/vm_param.h>
  83 #include <vm/pmap.h>
  84 #include <vm/vm_extern.h>
  85 #include <vm/vm_map.h>
  86 #include <vm/vm_kern.h>
  87 #include <vm/vm_object.h>
  88 #include <vm/vm_page.h>
  89 #include <vm/vm_pageout.h>
  90 #include <vm/vm_pager.h>
  91 #include <vm/swap_pager.h>
  92
  93 struct shm_mapping {
  94         char            *sm_path;
  95         Fnv32_t         sm_fnv;
  96         struct shmfd    *sm_shmfd;
  97         LIST_ENTRY(shm_mapping) sm_link;
  98 };
  99
 100 static MALLOC_DEFINE(M_SHMFD, "shmfd", "shared memory file descriptor");
 101 static LIST_HEAD(, shm_mapping) *shm_dictionary;
 102 static struct sx shm_dict_lock;
 103 static struct mtx shm_timestamp_lock;
 104 static u_long shm_hash;
 105 static struct unrhdr *shm_ino_unr;
 106 static dev_t shm_dev_ino;
 107
 108 #define SHM_HASH(fnv)   (&shm_dictionary[(fnv) & shm_hash])
 109
 110 static int      shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags);
 111 static struct shmfd *shm_alloc(struct ucred *ucred, mode_t mode);
 112 static void     shm_init(void *arg);
 113 static void     shm_drop(struct shmfd *shmfd);
 114 static struct shmfd *shm_hold(struct shmfd *shmfd);
 115 static void     shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd);
 116 static struct shmfd *shm_lookup(char *path, Fnv32_t fnv);
 117 static int      shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred);
 118 static int      shm_dotruncate(struct shmfd *shmfd, off_t length);
 119
 120 static fo_rdwr_t        shm_read;
 121 static fo_rdwr_t        shm_write;
 122 static fo_truncate_t    shm_truncate;
 123 static fo_ioctl_t       shm_ioctl;
 124 static fo_poll_t        shm_poll;
 125 static fo_kqfilter_t    shm_kqfilter;
 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
 132 /* File descriptor operations. */
 133 static struct fileops shm_ops = {
 134         .fo_read = shm_read,
 135         .fo_write = shm_write,
 136         .fo_truncate = shm_truncate,
 137         .fo_ioctl = shm_ioctl,
 138         .fo_poll = shm_poll,
 139         .fo_kqfilter = shm_kqfilter,
 140         .fo_stat = shm_stat,
 141         .fo_close = shm_close,
 142         .fo_chmod = shm_chmod,
 143         .fo_chown = shm_chown,
 144         .fo_sendfile = vn_sendfile,
 145         .fo_seek = shm_seek,
 146         .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
 147 };
 148
 149 FEATURE(posix_shm, "POSIX shared memory");
 150
 151 static int
 152 uiomove_object_page(vm_object_t obj, size_t len, struct uio *uio)
 153 {
 154         vm_page_t m;
 155         vm_pindex_t idx;
 156         size_t tlen;
 157         int error, offset, rv;
 158
 159         idx = OFF_TO_IDX(uio->uio_offset);
 160         offset = uio->uio_offset & PAGE_MASK;
 161         tlen = MIN(PAGE_SIZE - offset, len);
 162
 163         VM_OBJECT_WLOCK(obj);
 164
 165         /*
 166          * Parallel reads of the page content from disk are prevented
 167          * by exclusive busy.
 168          *
 169          * Although the tmpfs vnode lock is held here, it is
 170          * nonetheless safe to sleep waiting for a free page.  The
 171          * pageout daemon does not need to acquire the tmpfs vnode
 172          * lock to page out tobj's pages because tobj is a OBJT_SWAP
 173          * type object.
 174          */
 175         m = vm_page_grab(obj, idx, VM_ALLOC_NORMAL);
 176         if (m->valid != VM_PAGE_BITS_ALL) {
 177                 if (vm_pager_has_page(obj, idx, NULL, NULL)) {
 178                         rv = vm_pager_get_pages(obj, &m, 1, 0);
 179                         m = vm_page_lookup(obj, idx);
 180                         if (m == NULL) {
 181                                 printf(
 182                     "uiomove_object: vm_obj %p idx %jd null lookup rv %d\n",
 183                                     obj, idx, rv);
 184                                 VM_OBJECT_WUNLOCK(obj);
 185                                 return (EIO);
 186                         }
 187                         if (rv != VM_PAGER_OK) {
 188                                 printf(
 189             "uiomove_object: vm_obj %p idx %jd valid %x pager error %d\n",
 190                                     obj, idx, m->valid, rv);
 191                                 vm_page_lock(m);
 192                                 vm_page_free(m);
 193                                 vm_page_unlock(m);
 194                                 VM_OBJECT_WUNLOCK(obj);
 195                                 return (EIO);
 196                         }
 197                 } else
 198                         vm_page_zero_invalid(m, TRUE);
 199         }
 200         vm_page_xunbusy(m);
 201         vm_page_lock(m);
 202         vm_page_hold(m);
 203         if (m->queue == PQ_NONE) {
 204                 vm_page_deactivate(m);
 205         } else {
 206                 /* Requeue to maintain LRU ordering. */
 207                 vm_page_requeue(m);
 208         }
 209         vm_page_unlock(m);
 210         VM_OBJECT_WUNLOCK(obj);
 211         error = uiomove_fromphys(&m, offset, tlen, uio);
 212         if (uio->uio_rw == UIO_WRITE && error == 0) {
 213                 VM_OBJECT_WLOCK(obj);
 214                 vm_page_dirty(m);
 215                 vm_pager_page_unswapped(m);
 216                 VM_OBJECT_WUNLOCK(obj);
 217         }
 218         vm_page_lock(m);
 219         vm_page_unhold(m);
 220         vm_page_unlock(m);
 221
 222         return (error);
 223 }
 224
 225 int
 226 uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio)
 227 {
 228         ssize_t resid;
 229         size_t len;
 230         int error;
 231
 232         error = 0;
 233         while ((resid = uio->uio_resid) > 0) {
 234                 if (obj_size <= uio->uio_offset)
 235                         break;
 236                 len = MIN(obj_size - uio->uio_offset, resid);
 237                 if (len == 0)
 238                         break;
 239                 error = uiomove_object_page(obj, len, uio);
 240                 if (error != 0 || resid == uio->uio_resid)
 241                         break;
 242         }
 243         return (error);
 244 }
 245
 246 static int
 247 shm_seek(struct file *fp, off_t offset, int whence, struct thread *td)
 248 {
 249         struct shmfd *shmfd;
 250         off_t foffset;
 251         int error;
 252
 253         shmfd = fp->f_data;
 254         foffset = foffset_lock(fp, 0);
 255         error = 0;
 256         switch (whence) {
 257         case L_INCR:
 258                 if (foffset < 0 ||
 259                     (offset > 0 && foffset > OFF_MAX - offset)) {
 260                         error = EOVERFLOW;
 261                         break;
 262                 }
 263                 offset += foffset;
 264                 break;
 265         case L_XTND:
 266                 if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) {
 267                         error = EOVERFLOW;
 268                         break;
 269                 }
 270                 offset += shmfd->shm_size;
 271                 break;
 272         case L_SET:
 273                 break;
 274         default:
 275                 error = EINVAL;
 276         }
 277         if (error == 0) {
 278                 if (offset < 0 || offset > shmfd->shm_size)
 279                         error = EINVAL;
 280                 else
 281                         *(off_t *)(td->td_retval) = offset;
 282         }
 283         foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
 284         return (error);
 285 }
 286
 287 static int
 288 shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
 289     int flags, struct thread *td)
 290 {
 291         struct shmfd *shmfd;
 292         void *rl_cookie;
 293         int error;
 294
 295         shmfd = fp->f_data;
 296         foffset_lock_uio(fp, uio, flags);
 297         rl_cookie = rangelock_rlock(&shmfd->shm_rl, uio->uio_offset,
 298             uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
 299 #ifdef MAC
 300         error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd);
 301         if (error)
 302                 return (error);
 303 #endif
 304         error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
 305         rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
 306         foffset_unlock_uio(fp, uio, flags);
 307         return (error);
 308 }
 309
 310 static int
 311 shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
 312     int flags, struct thread *td)
 313 {
 314         struct shmfd *shmfd;
 315         void *rl_cookie;
 316         int error;
 317
 318         shmfd = fp->f_data;
 319 #ifdef MAC
 320         error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd);
 321         if (error)
 322                 return (error);
 323 #endif
 324         foffset_lock_uio(fp, uio, flags);
 325         if ((flags & FOF_OFFSET) == 0) {
 326                 rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
 327                     &shmfd->shm_mtx);
 328         } else {
 329                 rl_cookie = rangelock_wlock(&shmfd->shm_rl, uio->uio_offset,
 330                     uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
 331         }
 332
 333         error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
 334         rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
 335         foffset_unlock_uio(fp, uio, flags);
 336         return (error);
 337 }
 338
 339 static int
 340 shm_truncate(struct file *fp, off_t length, struct ucred *active_cred,
 341     struct thread *td)
 342 {
 343         struct shmfd *shmfd;
 344 #ifdef MAC
 345         int error;
 346 #endif
 347
 348         shmfd = fp->f_data;
 349 #ifdef MAC
 350         error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd);
 351         if (error)
 352                 return (error);
 353 #endif
 354         return (shm_dotruncate(shmfd, length));
 355 }
 356
 357 static int
 358 shm_ioctl(struct file *fp, u_long com, void *data,
 359     struct ucred *active_cred, struct thread *td)
 360 {
 361
 362         return (EOPNOTSUPP);
 363 }
 364
 365 static int
 366 shm_poll(struct file *fp, int events, struct ucred *active_cred,
 367     struct thread *td)
 368 {
 369
 370         return (EOPNOTSUPP);
 371 }
 372
 373 static int
 374 shm_kqfilter(struct file *fp, struct knote *kn)
 375 {
 376
 377         return (EOPNOTSUPP);
 378 }
 379
 380 static int
 381 shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred,
 382     struct thread *td)
 383 {
 384         struct shmfd *shmfd;
 385 #ifdef MAC
 386         int error;
 387 #endif
 388
 389         shmfd = fp->f_data;
 390
 391 #ifdef MAC
 392         error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd);
 393         if (error)
 394                 return (error);
 395 #endif
 396
 397         /*
 398          * Attempt to return sanish values for fstat() on a memory file
 399          * descriptor.
 400          */
 401         bzero(sb, sizeof(*sb));
 402         sb->st_blksize = PAGE_SIZE;
 403         sb->st_size = shmfd->shm_size;
 404         sb->st_blocks = (sb->st_size + sb->st_blksize - 1) / sb->st_blksize;
 405         mtx_lock(&shm_timestamp_lock);
 406         sb->st_atim = shmfd->shm_atime;
 407         sb->st_ctim = shmfd->shm_ctime;
 408         sb->st_mtim = shmfd->shm_mtime;
 409         sb->st_birthtim = shmfd->shm_birthtime;
 410         sb->st_mode = S_IFREG | shmfd->shm_mode;                /* XXX */
 411         sb->st_uid = shmfd->shm_uid;
 412         sb->st_gid = shmfd->shm_gid;
 413         mtx_unlock(&shm_timestamp_lock);
 414         sb->st_dev = shm_dev_ino;
 415         sb->st_ino = shmfd->shm_ino;
 416
 417         return (0);
 418 }
 419
 420 static int
 421 shm_close(struct file *fp, struct thread *td)
 422 {
 423         struct shmfd *shmfd;
 424
 425         shmfd = fp->f_data;
 426         fp->f_data = NULL;
 427         shm_drop(shmfd);
 428
 429         return (0);
 430 }
 431
 432 static int
 433 shm_dotruncate(struct shmfd *shmfd, off_t length)
 434 {
 435         vm_object_t object;
 436         vm_page_t m, ma[1];
 437         vm_pindex_t idx, nobjsize;
 438         vm_ooffset_t delta;
 439         int base, rv;
 440
 441         object = shmfd->shm_object;
 442         VM_OBJECT_WLOCK(object);
 443         if (length == shmfd->shm_size) {
 444                 VM_OBJECT_WUNLOCK(object);
 445                 return (0);
 446         }
 447         nobjsize = OFF_TO_IDX(length + PAGE_MASK);
 448
 449         /* Are we shrinking?  If so, trim the end. */
 450         if (length < shmfd->shm_size) {
 451                 /*
 452                  * Disallow any requests to shrink the size if this
 453                  * object is mapped into the kernel.
 454                  */
 455                 if (shmfd->shm_kmappings > 0) {
 456                         VM_OBJECT_WUNLOCK(object);
 457                         return (EBUSY);
 458                 }
 459
 460                 /*
 461                  * Zero the truncated part of the last page.
 462                  */
 463                 base = length & PAGE_MASK;
 464                 if (base != 0) {
 465                         idx = OFF_TO_IDX(length);
 466 retry:
 467                         m = vm_page_lookup(object, idx);
 468                         if (m != NULL) {
 469                                 if (vm_page_sleep_if_busy(m, "shmtrc"))
 470                                         goto retry;
 471                         } else if (vm_pager_has_page(object, idx, NULL, NULL)) {
 472                                 m = vm_page_alloc(object, idx, VM_ALLOC_NORMAL);
 473                                 if (m == NULL) {
 474                                         VM_OBJECT_WUNLOCK(object);
 475                                         VM_WAIT;
 476                                         VM_OBJECT_WLOCK(object);
 477                                         goto retry;
 478                                 } else if (m->valid != VM_PAGE_BITS_ALL) {
 479                                         ma[0] = m;
 480                                         rv = vm_pager_get_pages(object, ma, 1,
 481                                             0);
 482                                         m = vm_page_lookup(object, idx);
 483                                 } else
 484                                         /* A cached page was reactivated. */
 485                                         rv = VM_PAGER_OK;
 486                                 vm_page_lock(m);
 487                                 if (rv == VM_PAGER_OK) {
 488                                         vm_page_deactivate(m);
 489                                         vm_page_unlock(m);
 490                                         vm_page_xunbusy(m);
 491                                 } else {
 492                                         vm_page_free(m);
 493                                         vm_page_unlock(m);
 494                                         VM_OBJECT_WUNLOCK(object);
 495                                         return (EIO);
 496                                 }
 497                         }
 498                         if (m != NULL) {
 499                                 pmap_zero_page_area(m, base, PAGE_SIZE - base);
 500                                 KASSERT(m->valid == VM_PAGE_BITS_ALL,
 501                                     ("shm_dotruncate: page %p is invalid", m));
 502                                 vm_page_dirty(m);
 503                                 vm_pager_page_unswapped(m);
 504                         }
 505                 }
 506                 delta = ptoa(object->size - nobjsize);
 507
 508                 /* Toss in memory pages. */
 509                 if (nobjsize < object->size)
 510                         vm_object_page_remove(object, nobjsize, object->size,
 511                             0);
 512
 513                 /* Toss pages from swap. */
 514                 if (object->type == OBJT_SWAP)
 515                         swap_pager_freespace(object, nobjsize, delta);
 516
 517                 /* Free the swap accounted for shm */
 518                 swap_release_by_cred(delta, object->cred);
 519                 object->charge -= delta;
 520         } else {
 521                 /* Attempt to reserve the swap */
 522                 delta = ptoa(nobjsize - object->size);
 523                 if (!swap_reserve_by_cred(delta, object->cred)) {
 524                         VM_OBJECT_WUNLOCK(object);
 525                         return (ENOMEM);
 526                 }
 527                 object->charge += delta;
 528         }
 529         shmfd->shm_size = length;
 530         mtx_lock(&shm_timestamp_lock);
 531         vfs_timestamp(&shmfd->shm_ctime);
 532         shmfd->shm_mtime = shmfd->shm_ctime;
 533         mtx_unlock(&shm_timestamp_lock);
 534         object->size = nobjsize;
 535         VM_OBJECT_WUNLOCK(object);
 536         return (0);
 537 }
 538
 539 /*
 540  * shmfd object management including creation and reference counting
 541  * routines.
 542  */
 543 static struct shmfd *
 544 shm_alloc(struct ucred *ucred, mode_t mode)
 545 {
 546         struct shmfd *shmfd;
 547         int ino;
 548
 549         shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO);
 550         shmfd->shm_size = 0;
 551         shmfd->shm_uid = ucred->cr_uid;
 552         shmfd->shm_gid = ucred->cr_gid;
 553         shmfd->shm_mode = mode;
 554         shmfd->shm_object = vm_pager_allocate(OBJT_DEFAULT, NULL,
 555             shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred);
 556         KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate"));
 557         VM_OBJECT_WLOCK(shmfd->shm_object);
 558         vm_object_clear_flag(shmfd->shm_object, OBJ_ONEMAPPING);
 559         vm_object_set_flag(shmfd->shm_object, OBJ_NOSPLIT);
 560         VM_OBJECT_WUNLOCK(shmfd->shm_object);
 561         vfs_timestamp(&shmfd->shm_birthtime);
 562         shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime =
 563             shmfd->shm_birthtime;
 564         ino = alloc_unr(shm_ino_unr);
 565         if (ino == -1)
 566                 shmfd->shm_ino = 0;
 567         else
 568                 shmfd->shm_ino = ino;
 569         refcount_init(&shmfd->shm_refs, 1);
 570         mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF);
 571         rangelock_init(&shmfd->shm_rl);
 572 #ifdef MAC
 573         mac_posixshm_init(shmfd);
 574         mac_posixshm_create(ucred, shmfd);
 575 #endif
 576
 577         return (shmfd);
 578 }
 579
 580 static struct shmfd *
 581 shm_hold(struct shmfd *shmfd)
 582 {
 583
 584         refcount_acquire(&shmfd->shm_refs);
 585         return (shmfd);
 586 }
 587
 588 static void
 589 shm_drop(struct shmfd *shmfd)
 590 {
 591
 592         if (refcount_release(&shmfd->shm_refs)) {
 593 #ifdef MAC
 594                 mac_posixshm_destroy(shmfd);
 595 #endif
 596                 rangelock_destroy(&shmfd->shm_rl);
 597                 mtx_destroy(&shmfd->shm_mtx);
 598                 vm_object_deallocate(shmfd->shm_object);
 599                 if (shmfd->shm_ino != 0)
 600                         free_unr(shm_ino_unr, shmfd->shm_ino);
 601                 free(shmfd, M_SHMFD);
 602         }
 603 }
 604
 605 /*
 606  * Determine if the credentials have sufficient permissions for a
 607  * specified combination of FREAD and FWRITE.
 608  */
 609 static int
 610 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags)
 611 {
 612         accmode_t accmode;
 613         int error;
 614
 615         accmode = 0;
 616         if (flags & FREAD)
 617                 accmode |= VREAD;
 618         if (flags & FWRITE)
 619                 accmode |= VWRITE;
 620         mtx_lock(&shm_timestamp_lock);
 621         error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
 622             accmode, ucred, NULL);
 623         mtx_unlock(&shm_timestamp_lock);
 624         return (error);
 625 }
 626
 627 /*
 628  * Dictionary management.  We maintain an in-kernel dictionary to map
 629  * paths to shmfd objects.  We use the FNV hash on the path to store
 630  * the mappings in a hash table.
 631  */
 632 static void
 633 shm_init(void *arg)
 634 {
 635
 636         mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF);
 637         sx_init(&shm_dict_lock, "shm dictionary");
 638         shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash);
 639         shm_ino_unr = new_unrhdr(1, INT32_MAX, NULL);
 640         KASSERT(shm_ino_unr != NULL, ("shm fake inodes not initialized"));
 641         shm_dev_ino = devfs_alloc_cdp_inode();
 642         KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized"));
 643 }
 644 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL);
 645
 646 static struct shmfd *
 647 shm_lookup(char *path, Fnv32_t fnv)
 648 {
 649         struct shm_mapping *map;
 650
 651         LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
 652                 if (map->sm_fnv != fnv)
 653                         continue;
 654                 if (strcmp(map->sm_path, path) == 0)
 655                         return (map->sm_shmfd);
 656         }
 657
 658         return (NULL);
 659 }
 660
 661 static void
 662 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd)
 663 {
 664         struct shm_mapping *map;
 665
 666         map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK);
 667         map->sm_path = path;
 668         map->sm_fnv = fnv;
 669         map->sm_shmfd = shm_hold(shmfd);
 670         shmfd->shm_path = path;
 671         LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link);
 672 }
 673
 674 static int
 675 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred)
 676 {
 677         struct shm_mapping *map;
 678         int error;
 679
 680         LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
 681                 if (map->sm_fnv != fnv)
 682                         continue;
 683                 if (strcmp(map->sm_path, path) == 0) {
 684 #ifdef MAC
 685                         error = mac_posixshm_check_unlink(ucred, map->sm_shmfd);
 686                         if (error)
 687                                 return (error);
 688 #endif
 689                         error = shm_access(map->sm_shmfd, ucred,
 690                             FREAD | FWRITE);
 691                         if (error)
 692                                 return (error);
 693                         map->sm_shmfd->shm_path = NULL;
 694                         LIST_REMOVE(map, sm_link);
 695                         shm_drop(map->sm_shmfd);
 696                         free(map->sm_path, M_SHMFD);
 697                         free(map, M_SHMFD);
 698                         return (0);
 699                 }
 700         }
 701
 702         return (ENOENT);
 703 }
 704
 705 /* System calls. */
 706 int
 707 sys_shm_open(struct thread *td, struct shm_open_args *uap)
 708 {
 709         struct filedesc *fdp;
 710         struct shmfd *shmfd;
 711         struct file *fp;
 712         char *path;
 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) && (uap->path != SHM_ANON))
 722                 return (ECAPMODE);
 723 #endif
 724
 725         if ((uap->flags & O_ACCMODE) != O_RDONLY &&
 726             (uap->flags & O_ACCMODE) != O_RDWR)
 727                 return (EINVAL);
 728
 729         if ((uap->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 = (uap->mode & ~fdp->fd_cmask) & ACCESSPERMS;
 734
 735         error = falloc(td, &fp, &fd, O_CLOEXEC);
 736         if (error)
 737                 return (error);
 738
 739         /* A SHM_ANON path pointer creates an anonymous object. */
 740         if (uap->path == SHM_ANON) {
 741                 /* A read-only anonymous object is pointless. */
 742                 if ((uap->flags & O_ACCMODE) == O_RDONLY) {
 743                         fdclose(fdp, fp, fd, td);
 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                 error = copyinstr(uap->path, path, MAXPATHLEN, NULL);
 751 #ifdef KTRACE
 752                 if (error == 0 && KTRPOINT(curthread, KTR_NAMEI))
 753                         ktrnamei(path);
 754 #endif
 755                 /* Require paths to start with a '/' character. */
 756                 if (error == 0 && path[0] != '/')
 757                         error = EINVAL;
 758                 if (error) {
 759                         fdclose(fdp, fp, fd, td);
 760                         fdrop(fp, td);
 761                         free(path, M_SHMFD);
 762                         return (error);
 763                 }
 764
 765                 fnv = fnv_32_str(path, FNV1_32_INIT);
 766                 sx_xlock(&shm_dict_lock);
 767                 shmfd = shm_lookup(path, fnv);
 768                 if (shmfd == NULL) {
 769                         /* Object does not yet exist, create it if requested. */
 770                         if (uap->flags & O_CREAT) {
 771 #ifdef MAC
 772                                 error = mac_posixshm_check_create(td->td_ucred,
 773                                     path);
 774                                 if (error == 0) {
 775 #endif
 776                                         shmfd = shm_alloc(td->td_ucred, cmode);
 777                                         shm_insert(path, fnv, shmfd);
 778 #ifdef MAC
 779                                 }
 780 #endif
 781                         } else {
 782                                 free(path, M_SHMFD);
 783                                 error = ENOENT;
 784                         }
 785                 } else {
 786                         /*
 787                          * Object already exists, obtain a new
 788                          * reference if requested and permitted.
 789                          */
 790                         free(path, M_SHMFD);
 791                         if ((uap->flags & (O_CREAT | O_EXCL)) ==
 792                             (O_CREAT | O_EXCL))
 793                                 error = EEXIST;
 794                         else {
 795 #ifdef MAC
 796                                 error = mac_posixshm_check_open(td->td_ucred,
 797                                     shmfd, FFLAGS(uap->flags & O_ACCMODE));
 798                                 if (error == 0)
 799 #endif
 800                                 error = shm_access(shmfd, td->td_ucred,
 801                                     FFLAGS(uap->flags & O_ACCMODE));
 802                         }
 803
 804                         /*
 805                          * Truncate the file back to zero length if
 806                          * O_TRUNC was specified and the object was
 807                          * opened with read/write.
 808                          */
 809                         if (error == 0 &&
 810                             (uap->flags & (O_ACCMODE | O_TRUNC)) ==
 811                             (O_RDWR | O_TRUNC)) {
 812 #ifdef MAC
 813                                 error = mac_posixshm_check_truncate(
 814                                         td->td_ucred, fp->f_cred, shmfd);
 815                                 if (error == 0)
 816 #endif
 817                                         shm_dotruncate(shmfd, 0);
 818                         }
 819                         if (error == 0)
 820                                 shm_hold(shmfd);
 821                 }
 822                 sx_xunlock(&shm_dict_lock);
 823
 824                 if (error) {
 825                         fdclose(fdp, fp, fd, td);
 826                         fdrop(fp, td);
 827                         return (error);
 828                 }
 829         }
 830
 831         finit(fp, FFLAGS(uap->flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops);
 832
 833         td->td_retval[0] = fd;
 834         fdrop(fp, td);
 835
 836         return (0);
 837 }
 838
 839 int
 840 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap)
 841 {
 842         char *path;
 843         Fnv32_t fnv;
 844         int error;
 845
 846         path = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
 847         error = copyinstr(uap->path, path, MAXPATHLEN, NULL);
 848         if (error) {
 849                 free(path, M_TEMP);
 850                 return (error);
 851         }
 852 #ifdef KTRACE
 853         if (KTRPOINT(curthread, KTR_NAMEI))
 854                 ktrnamei(path);
 855 #endif
 856         fnv = fnv_32_str(path, FNV1_32_INIT);
 857         sx_xlock(&shm_dict_lock);
 858         error = shm_remove(path, fnv, td->td_ucred);
 859         sx_xunlock(&shm_dict_lock);
 860         free(path, M_TEMP);
 861
 862         return (error);
 863 }
 864
 865 /*
 866  * mmap() helper to validate mmap() requests against shm object state
 867  * and give mmap() the vm_object to use for the mapping.
 868  */
 869 int
 870 shm_mmap(struct shmfd *shmfd, vm_size_t objsize, vm_ooffset_t foff,
 871     vm_object_t *obj)
 872 {
 873
 874         /*
 875          * XXXRW: This validation is probably insufficient, and subject to
 876          * sign errors.  It should be fixed.
 877          */
 878         if (foff >= shmfd->shm_size ||
 879             foff + objsize > round_page(shmfd->shm_size))
 880                 return (EINVAL);
 881
 882         mtx_lock(&shm_timestamp_lock);
 883         vfs_timestamp(&shmfd->shm_atime);
 884         mtx_unlock(&shm_timestamp_lock);
 885         vm_object_reference(shmfd->shm_object);
 886         *obj = shmfd->shm_object;
 887         return (0);
 888 }
 889
 890 static int
 891 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
 892     struct thread *td)
 893 {
 894         struct shmfd *shmfd;
 895         int error;
 896
 897         error = 0;
 898         shmfd = fp->f_data;
 899         mtx_lock(&shm_timestamp_lock);
 900         /*
 901          * SUSv4 says that x bits of permission need not be affected.
 902          * Be consistent with our shm_open there.
 903          */
 904 #ifdef MAC
 905         error = mac_posixshm_check_setmode(active_cred, shmfd, mode);
 906         if (error != 0)
 907                 goto out;
 908 #endif
 909         error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid,
 910             shmfd->shm_gid, VADMIN, active_cred, NULL);
 911         if (error != 0)
 912                 goto out;
 913         shmfd->shm_mode = mode & ACCESSPERMS;
 914 out:
 915         mtx_unlock(&shm_timestamp_lock);
 916         return (error);
 917 }
 918
 919 static int
 920 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
 921     struct thread *td)
 922 {
 923         struct shmfd *shmfd;
 924         int error;
 925
 926         error = 0;
 927         shmfd = fp->f_data;
 928         mtx_lock(&shm_timestamp_lock);
 929 #ifdef MAC
 930         error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid);
 931         if (error != 0)
 932                 goto out;
 933 #endif
 934         if (uid == (uid_t)-1)
 935                 uid = shmfd->shm_uid;
 936         if (gid == (gid_t)-1)
 937                  gid = shmfd->shm_gid;
 938         if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) ||
 939             (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) &&
 940             (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN, 0)))
 941                 goto out;
 942         shmfd->shm_uid = uid;
 943         shmfd->shm_gid = gid;
 944 out:
 945         mtx_unlock(&shm_timestamp_lock);
 946         return (error);
 947 }
 948
 949 /*
 950  * Helper routines to allow the backing object of a shared memory file
 951  * descriptor to be mapped in the kernel.
 952  */
 953 int
 954 shm_map(struct file *fp, size_t size, off_t offset, void **memp)
 955 {
 956         struct shmfd *shmfd;
 957         vm_offset_t kva, ofs;
 958         vm_object_t obj;
 959         int rv;
 960
 961         if (fp->f_type != DTYPE_SHM)
 962                 return (EINVAL);
 963         shmfd = fp->f_data;
 964         obj = shmfd->shm_object;
 965         VM_OBJECT_WLOCK(obj);
 966         /*
 967          * XXXRW: This validation is probably insufficient, and subject to
 968          * sign errors.  It should be fixed.
 969          */
 970         if (offset >= shmfd->shm_size ||
 971             offset + size > round_page(shmfd->shm_size)) {
 972                 VM_OBJECT_WUNLOCK(obj);
 973                 return (EINVAL);
 974         }
 975
 976         shmfd->shm_kmappings++;
 977         vm_object_reference_locked(obj);
 978         VM_OBJECT_WUNLOCK(obj);
 979
 980         /* Map the object into the kernel_map and wire it. */
 981         kva = vm_map_min(kernel_map);
 982         ofs = offset & PAGE_MASK;
 983         offset = trunc_page(offset);
 984         size = round_page(size + ofs);
 985         rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0,
 986             VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE,
 987             VM_PROT_READ | VM_PROT_WRITE, 0);
 988         if (rv == KERN_SUCCESS) {
 989                 rv = vm_map_wire(kernel_map, kva, kva + size,
 990                     VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
 991                 if (rv == KERN_SUCCESS) {
 992                         *memp = (void *)(kva + ofs);
 993                         return (0);
 994                 }
 995                 vm_map_remove(kernel_map, kva, kva + size);
 996         } else
 997                 vm_object_deallocate(obj);
 998
 999         /* On failure, drop our mapping reference. */
1000         VM_OBJECT_WLOCK(obj);
1001         shmfd->shm_kmappings--;
1002         VM_OBJECT_WUNLOCK(obj);
1003
1004         return (vm_mmap_to_errno(rv));
1005 }
1006
1007 /*
1008  * We require the caller to unmap the entire entry.  This allows us to
1009  * safely decrement shm_kmappings when a mapping is removed.
1010  */
1011 int
1012 shm_unmap(struct file *fp, void *mem, size_t size)
1013 {
1014         struct shmfd *shmfd;
1015         vm_map_entry_t entry;
1016         vm_offset_t kva, ofs;
1017         vm_object_t obj;
1018         vm_pindex_t pindex;
1019         vm_prot_t prot;
1020         boolean_t wired;
1021         vm_map_t map;
1022         int rv;
1023
1024         if (fp->f_type != DTYPE_SHM)
1025                 return (EINVAL);
1026         shmfd = fp->f_data;
1027         kva = (vm_offset_t)mem;
1028         ofs = kva & PAGE_MASK;
1029         kva = trunc_page(kva);
1030         size = round_page(size + ofs);
1031         map = kernel_map;
1032         rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry,
1033             &obj, &pindex, &prot, &wired);
1034         if (rv != KERN_SUCCESS)
1035                 return (EINVAL);
1036         if (entry->start != kva || entry->end != kva + size) {
1037                 vm_map_lookup_done(map, entry);
1038                 return (EINVAL);
1039         }
1040         vm_map_lookup_done(map, entry);
1041         if (obj != shmfd->shm_object)
1042                 return (EINVAL);
1043         vm_map_remove(map, kva, kva + size);
1044         VM_OBJECT_WLOCK(obj);
1045         KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped"));
1046         shmfd->shm_kmappings--;
1047         VM_OBJECT_WUNLOCK(obj);
1048         return (0);
1049 }
1050
1051 void
1052 shm_path(struct shmfd *shmfd, char *path, size_t size)
1053 {
1054
1055         if (shmfd->shm_path == NULL)
1056                 return;
1057         sx_slock(&shm_dict_lock);
1058         if (shmfd->shm_path != NULL)
1059                 strlcpy(path, shmfd->shm_path, size);
1060         sx_sunlock(&shm_dict_lock);
1061 }