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/capsicum.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/jail.h>
  61 #include <sys/ktrace.h>
  62 #include <sys/lock.h>
  63 #include <sys/malloc.h>
  64 #include <sys/mman.h>
  65 #include <sys/mutex.h>
  66 #include <sys/priv.h>
  67 #include <sys/proc.h>
  68 #include <sys/refcount.h>
  69 #include <sys/resourcevar.h>
  70 #include <sys/rwlock.h>
  71 #include <sys/stat.h>
  72 #include <sys/sysctl.h>
  73 #include <sys/sysproto.h>
  74 #include <sys/systm.h>
  75 #include <sys/sx.h>
  76 #include <sys/time.h>
  77 #include <sys/vnode.h>
  78 #include <sys/unistd.h>
  79
  80 #include <security/mac/mac_framework.h>
  81
  82 #include <vm/vm.h>
  83 #include <vm/vm_param.h>
  84 #include <vm/pmap.h>
  85 #include <vm/vm_extern.h>
  86 #include <vm/vm_map.h>
  87 #include <vm/vm_kern.h>
  88 #include <vm/vm_object.h>
  89 #include <vm/vm_page.h>
  90 #include <vm/vm_pageout.h>
  91 #include <vm/vm_pager.h>
  92 #include <vm/swap_pager.h>
  93
  94 struct shm_mapping {
  95         char            *sm_path;
  96         Fnv32_t         sm_fnv;
  97         struct shmfd    *sm_shmfd;
  98         LIST_ENTRY(shm_mapping) sm_link;
  99 };
 100
 101 static MALLOC_DEFINE(M_SHMFD, "shmfd", "shared memory file descriptor");
 102 static LIST_HEAD(, shm_mapping) *shm_dictionary;
 103 static struct sx shm_dict_lock;
 104 static struct mtx shm_timestamp_lock;
 105 static u_long shm_hash;
 106 static struct unrhdr *shm_ino_unr;
 107 static dev_t shm_dev_ino;
 108
 109 #define SHM_HASH(fnv)   (&shm_dictionary[(fnv) & shm_hash])
 110
 111 static int      shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags);
 112 static struct shmfd *shm_alloc(struct ucred *ucred, mode_t mode);
 113 static void     shm_init(void *arg);
 114 static void     shm_drop(struct shmfd *shmfd);
 115 static struct shmfd *shm_hold(struct shmfd *shmfd);
 116 static void     shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd);
 117 static struct shmfd *shm_lookup(char *path, Fnv32_t fnv);
 118 static int      shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred);
 119 static int      shm_dotruncate(struct shmfd *shmfd, off_t length);
 120
 121 static fo_rdwr_t        shm_read;
 122 static fo_rdwr_t        shm_write;
 123 static fo_truncate_t    shm_truncate;
 124 static fo_ioctl_t       shm_ioctl;
 125 static fo_poll_t        shm_poll;
 126 static fo_kqfilter_t    shm_kqfilter;
 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
 133 /* File descriptor operations. */
 134 static struct fileops shm_ops = {
 135         .fo_read = shm_read,
 136         .fo_write = shm_write,
 137         .fo_truncate = shm_truncate,
 138         .fo_ioctl = shm_ioctl,
 139         .fo_poll = shm_poll,
 140         .fo_kqfilter = shm_kqfilter,
 141         .fo_stat = shm_stat,
 142         .fo_close = shm_close,
 143         .fo_chmod = shm_chmod,
 144         .fo_chown = shm_chown,
 145         .fo_sendfile = vn_sendfile,
 146         .fo_seek = shm_seek,
 147         .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
 148 };
 149
 150 FEATURE(posix_shm, "POSIX shared memory");
 151
 152 static int
 153 uiomove_object_page(vm_object_t obj, size_t len, struct uio *uio)
 154 {
 155         vm_page_t m;
 156         vm_pindex_t idx;
 157         size_t tlen;
 158         int error, offset, rv;
 159
 160         idx = OFF_TO_IDX(uio->uio_offset);
 161         offset = uio->uio_offset & PAGE_MASK;
 162         tlen = MIN(PAGE_SIZE - offset, len);
 163
 164         VM_OBJECT_WLOCK(obj);
 165
 166         /*
 167          * Parallel reads of the page content from disk are prevented
 168          * by exclusive busy.
 169          *
 170          * Although the tmpfs vnode lock is held here, it is
 171          * nonetheless safe to sleep waiting for a free page.  The
 172          * pageout daemon does not need to acquire the tmpfs vnode
 173          * lock to page out tobj's pages because tobj is a OBJT_SWAP
 174          * type object.
 175          */
 176         m = vm_page_grab(obj, idx, VM_ALLOC_NORMAL);
 177         if (m->valid != VM_PAGE_BITS_ALL) {
 178                 if (vm_pager_has_page(obj, idx, NULL, NULL)) {
 179                         rv = vm_pager_get_pages(obj, &m, 1, 0);
 180                         m = vm_page_lookup(obj, idx);
 181                         if (m == NULL) {
 182                                 printf(
 183                     "uiomove_object: vm_obj %p idx %jd null lookup rv %d\n",
 184                                     obj, idx, rv);
 185                                 VM_OBJECT_WUNLOCK(obj);
 186                                 return (EIO);
 187                         }
 188                         if (rv != VM_PAGER_OK) {
 189                                 printf(
 190             "uiomove_object: vm_obj %p idx %jd valid %x pager error %d\n",
 191                                     obj, idx, m->valid, rv);
 192                                 vm_page_lock(m);
 193                                 vm_page_free(m);
 194                                 vm_page_unlock(m);
 195                                 VM_OBJECT_WUNLOCK(obj);
 196                                 return (EIO);
 197                         }
 198                 } else
 199                         vm_page_zero_invalid(m, TRUE);
 200         }
 201         vm_page_xunbusy(m);
 202         vm_page_lock(m);
 203         vm_page_hold(m);
 204         if (m->queue == PQ_NONE) {
 205                 vm_page_deactivate(m);
 206         } else {
 207                 /* Requeue to maintain LRU ordering. */
 208                 vm_page_requeue(m);
 209         }
 210         vm_page_unlock(m);
 211         VM_OBJECT_WUNLOCK(obj);
 212         error = uiomove_fromphys(&m, offset, tlen, uio);
 213         if (uio->uio_rw == UIO_WRITE && error == 0) {
 214                 VM_OBJECT_WLOCK(obj);
 215                 vm_page_dirty(m);
 216                 vm_pager_page_unswapped(m);
 217                 VM_OBJECT_WUNLOCK(obj);
 218         }
 219         vm_page_lock(m);
 220         vm_page_unhold(m);
 221         vm_page_unlock(m);
 222
 223         return (error);
 224 }
 225
 226 int
 227 uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio)
 228 {
 229         ssize_t resid;
 230         size_t len;
 231         int error;
 232
 233         error = 0;
 234         while ((resid = uio->uio_resid) > 0) {
 235                 if (obj_size <= uio->uio_offset)
 236                         break;
 237                 len = MIN(obj_size - uio->uio_offset, resid);
 238                 if (len == 0)
 239                         break;
 240                 error = uiomove_object_page(obj, len, uio);
 241                 if (error != 0 || resid == uio->uio_resid)
 242                         break;
 243         }
 244         return (error);
 245 }
 246
 247 static int
 248 shm_seek(struct file *fp, off_t offset, int whence, struct thread *td)
 249 {
 250         struct shmfd *shmfd;
 251         off_t foffset;
 252         int error;
 253
 254         shmfd = fp->f_data;
 255         foffset = foffset_lock(fp, 0);
 256         error = 0;
 257         switch (whence) {
 258         case L_INCR:
 259                 if (foffset < 0 ||
 260                     (offset > 0 && foffset > OFF_MAX - offset)) {
 261                         error = EOVERFLOW;
 262                         break;
 263                 }
 264                 offset += foffset;
 265                 break;
 266         case L_XTND:
 267                 if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) {
 268                         error = EOVERFLOW;
 269                         break;
 270                 }
 271                 offset += shmfd->shm_size;
 272                 break;
 273         case L_SET:
 274                 break;
 275         default:
 276                 error = EINVAL;
 277         }
 278         if (error == 0) {
 279                 if (offset < 0 || offset > shmfd->shm_size)
 280                         error = EINVAL;
 281                 else
 282                         *(off_t *)(td->td_retval) = offset;
 283         }
 284         foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
 285         return (error);
 286 }
 287
 288 static int
 289 shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
 290     int flags, struct thread *td)
 291 {
 292         struct shmfd *shmfd;
 293         void *rl_cookie;
 294         int error;
 295
 296         shmfd = fp->f_data;
 297         foffset_lock_uio(fp, uio, flags);
 298         rl_cookie = rangelock_rlock(&shmfd->shm_rl, uio->uio_offset,
 299             uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
 300 #ifdef MAC
 301         error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd);
 302         if (error)
 303                 return (error);
 304 #endif
 305         error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
 306         rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
 307         foffset_unlock_uio(fp, uio, flags);
 308         return (error);
 309 }
 310
 311 static int
 312 shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
 313     int flags, struct thread *td)
 314 {
 315         struct shmfd *shmfd;
 316         void *rl_cookie;
 317         int error;
 318
 319         shmfd = fp->f_data;
 320 #ifdef MAC
 321         error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd);
 322         if (error)
 323                 return (error);
 324 #endif
 325         foffset_lock_uio(fp, uio, flags);
 326         if ((flags & FOF_OFFSET) == 0) {
 327                 rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
 328                     &shmfd->shm_mtx);
 329         } else {
 330                 rl_cookie = rangelock_wlock(&shmfd->shm_rl, uio->uio_offset,
 331                     uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
 332         }
 333
 334         error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
 335         rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
 336         foffset_unlock_uio(fp, uio, flags);
 337         return (error);
 338 }
 339
 340 static int
 341 shm_truncate(struct file *fp, off_t length, struct ucred *active_cred,
 342     struct thread *td)
 343 {
 344         struct shmfd *shmfd;
 345 #ifdef MAC
 346         int error;
 347 #endif
 348
 349         shmfd = fp->f_data;
 350 #ifdef MAC
 351         error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd);
 352         if (error)
 353                 return (error);
 354 #endif
 355         return (shm_dotruncate(shmfd, length));
 356 }
 357
 358 static int
 359 shm_ioctl(struct file *fp, u_long com, void *data,
 360     struct ucred *active_cred, struct thread *td)
 361 {
 362
 363         return (EOPNOTSUPP);
 364 }
 365
 366 static int
 367 shm_poll(struct file *fp, int events, struct ucred *active_cred,
 368     struct thread *td)
 369 {
 370
 371         return (EOPNOTSUPP);
 372 }
 373
 374 static int
 375 shm_kqfilter(struct file *fp, struct knote *kn)
 376 {
 377
 378         return (EOPNOTSUPP);
 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 = (sb->st_size + sb->st_blksize - 1) / 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
 418         return (0);
 419 }
 420
 421 static int
 422 shm_close(struct file *fp, struct thread *td)
 423 {
 424         struct shmfd *shmfd;
 425
 426         shmfd = fp->f_data;
 427         fp->f_data = NULL;
 428         shm_drop(shmfd);
 429
 430         return (0);
 431 }
 432
 433 static int
 434 shm_dotruncate(struct shmfd *shmfd, off_t length)
 435 {
 436         vm_object_t object;
 437         vm_page_t m, ma[1];
 438         vm_pindex_t idx, nobjsize;
 439         vm_ooffset_t delta;
 440         int base, rv;
 441
 442         object = shmfd->shm_object;
 443         VM_OBJECT_WLOCK(object);
 444         if (length == shmfd->shm_size) {
 445                 VM_OBJECT_WUNLOCK(object);
 446                 return (0);
 447         }
 448         nobjsize = OFF_TO_IDX(length + PAGE_MASK);
 449
 450         /* Are we shrinking?  If so, trim the end. */
 451         if (length < shmfd->shm_size) {
 452                 /*
 453                  * Disallow any requests to shrink the size if this
 454                  * object is mapped into the kernel.
 455                  */
 456                 if (shmfd->shm_kmappings > 0) {
 457                         VM_OBJECT_WUNLOCK(object);
 458                         return (EBUSY);
 459                 }
 460
 461                 /*
 462                  * Zero the truncated part of the last page.
 463                  */
 464                 base = length & PAGE_MASK;
 465                 if (base != 0) {
 466                         idx = OFF_TO_IDX(length);
 467 retry:
 468                         m = vm_page_lookup(object, idx);
 469                         if (m != NULL) {
 470                                 if (vm_page_sleep_if_busy(m, "shmtrc"))
 471                                         goto retry;
 472                         } else if (vm_pager_has_page(object, idx, NULL, NULL)) {
 473                                 m = vm_page_alloc(object, idx, VM_ALLOC_NORMAL);
 474                                 if (m == NULL) {
 475                                         VM_OBJECT_WUNLOCK(object);
 476                                         VM_WAIT;
 477                                         VM_OBJECT_WLOCK(object);
 478                                         goto retry;
 479                                 } else if (m->valid != VM_PAGE_BITS_ALL) {
 480                                         ma[0] = m;
 481                                         rv = vm_pager_get_pages(object, ma, 1,
 482                                             0);
 483                                         m = vm_page_lookup(object, idx);
 484                                 } else
 485                                         /* A cached page was reactivated. */
 486                                         rv = VM_PAGER_OK;
 487                                 vm_page_lock(m);
 488                                 if (rv == VM_PAGER_OK) {
 489                                         vm_page_deactivate(m);
 490                                         vm_page_unlock(m);
 491                                         vm_page_xunbusy(m);
 492                                 } else {
 493                                         vm_page_free(m);
 494                                         vm_page_unlock(m);
 495                                         VM_OBJECT_WUNLOCK(object);
 496                                         return (EIO);
 497                                 }
 498                         }
 499                         if (m != NULL) {
 500                                 pmap_zero_page_area(m, base, PAGE_SIZE - base);
 501                                 KASSERT(m->valid == VM_PAGE_BITS_ALL,
 502                                     ("shm_dotruncate: page %p is invalid", m));
 503                                 vm_page_dirty(m);
 504                                 vm_pager_page_unswapped(m);
 505                         }
 506                 }
 507                 delta = ptoa(object->size - nobjsize);
 508
 509                 /* Toss in memory pages. */
 510                 if (nobjsize < object->size)
 511                         vm_object_page_remove(object, nobjsize, object->size,
 512                             0);
 513
 514                 /* Toss pages from swap. */
 515                 if (object->type == OBJT_SWAP)
 516                         swap_pager_freespace(object, nobjsize, delta);
 517
 518                 /* Free the swap accounted for shm */
 519                 swap_release_by_cred(delta, object->cred);
 520                 object->charge -= delta;
 521         } else {
 522                 /* Attempt to reserve the swap */
 523                 delta = ptoa(nobjsize - object->size);
 524                 if (!swap_reserve_by_cred(delta, object->cred)) {
 525                         VM_OBJECT_WUNLOCK(object);
 526                         return (ENOMEM);
 527                 }
 528                 object->charge += delta;
 529         }
 530         shmfd->shm_size = length;
 531         mtx_lock(&shm_timestamp_lock);
 532         vfs_timestamp(&shmfd->shm_ctime);
 533         shmfd->shm_mtime = shmfd->shm_ctime;
 534         mtx_unlock(&shm_timestamp_lock);
 535         object->size = nobjsize;
 536         VM_OBJECT_WUNLOCK(object);
 537         return (0);
 538 }
 539
 540 /*
 541  * shmfd object management including creation and reference counting
 542  * routines.
 543  */
 544 static struct shmfd *
 545 shm_alloc(struct ucred *ucred, mode_t mode)
 546 {
 547         struct shmfd *shmfd;
 548         int ino;
 549
 550         shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO);
 551         shmfd->shm_size = 0;
 552         shmfd->shm_uid = ucred->cr_uid;
 553         shmfd->shm_gid = ucred->cr_gid;
 554         shmfd->shm_mode = mode;
 555         shmfd->shm_object = vm_pager_allocate(OBJT_DEFAULT, NULL,
 556             shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred);
 557         KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate"));
 558         shmfd->shm_object->pg_color = 0;
 559         VM_OBJECT_WLOCK(shmfd->shm_object);
 560         vm_object_clear_flag(shmfd->shm_object, OBJ_ONEMAPPING);
 561         vm_object_set_flag(shmfd->shm_object, OBJ_COLORED | OBJ_NOSPLIT);
 562         VM_OBJECT_WUNLOCK(shmfd->shm_object);
 563         vfs_timestamp(&shmfd->shm_birthtime);
 564         shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime =
 565             shmfd->shm_birthtime;
 566         ino = alloc_unr(shm_ino_unr);
 567         if (ino == -1)
 568                 shmfd->shm_ino = 0;
 569         else
 570                 shmfd->shm_ino = ino;
 571         refcount_init(&shmfd->shm_refs, 1);
 572         mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF);
 573         rangelock_init(&shmfd->shm_rl);
 574 #ifdef MAC
 575         mac_posixshm_init(shmfd);
 576         mac_posixshm_create(ucred, shmfd);
 577 #endif
 578
 579         return (shmfd);
 580 }
 581
 582 static struct shmfd *
 583 shm_hold(struct shmfd *shmfd)
 584 {
 585
 586         refcount_acquire(&shmfd->shm_refs);
 587         return (shmfd);
 588 }
 589
 590 static void
 591 shm_drop(struct shmfd *shmfd)
 592 {
 593
 594         if (refcount_release(&shmfd->shm_refs)) {
 595 #ifdef MAC
 596                 mac_posixshm_destroy(shmfd);
 597 #endif
 598                 rangelock_destroy(&shmfd->shm_rl);
 599                 mtx_destroy(&shmfd->shm_mtx);
 600                 vm_object_deallocate(shmfd->shm_object);
 601                 if (shmfd->shm_ino != 0)
 602                         free_unr(shm_ino_unr, shmfd->shm_ino);
 603                 free(shmfd, M_SHMFD);
 604         }
 605 }
 606
 607 /*
 608  * Determine if the credentials have sufficient permissions for a
 609  * specified combination of FREAD and FWRITE.
 610  */
 611 static int
 612 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags)
 613 {
 614         accmode_t accmode;
 615         int error;
 616
 617         accmode = 0;
 618         if (flags & FREAD)
 619                 accmode |= VREAD;
 620         if (flags & FWRITE)
 621                 accmode |= VWRITE;
 622         mtx_lock(&shm_timestamp_lock);
 623         error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
 624             accmode, ucred, NULL);
 625         mtx_unlock(&shm_timestamp_lock);
 626         return (error);
 627 }
 628
 629 /*
 630  * Dictionary management.  We maintain an in-kernel dictionary to map
 631  * paths to shmfd objects.  We use the FNV hash on the path to store
 632  * the mappings in a hash table.
 633  */
 634 static void
 635 shm_init(void *arg)
 636 {
 637
 638         mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF);
 639         sx_init(&shm_dict_lock, "shm dictionary");
 640         shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash);
 641         shm_ino_unr = new_unrhdr(1, INT32_MAX, NULL);
 642         KASSERT(shm_ino_unr != NULL, ("shm fake inodes not initialized"));
 643         shm_dev_ino = devfs_alloc_cdp_inode();
 644         KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized"));
 645 }
 646 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL);
 647
 648 static struct shmfd *
 649 shm_lookup(char *path, Fnv32_t fnv)
 650 {
 651         struct shm_mapping *map;
 652
 653         LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
 654                 if (map->sm_fnv != fnv)
 655                         continue;
 656                 if (strcmp(map->sm_path, path) == 0)
 657                         return (map->sm_shmfd);
 658         }
 659
 660         return (NULL);
 661 }
 662
 663 static void
 664 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd)
 665 {
 666         struct shm_mapping *map;
 667
 668         map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK);
 669         map->sm_path = path;
 670         map->sm_fnv = fnv;
 671         map->sm_shmfd = shm_hold(shmfd);
 672         shmfd->shm_path = path;
 673         LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link);
 674 }
 675
 676 static int
 677 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred)
 678 {
 679         struct shm_mapping *map;
 680         int error;
 681
 682         LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
 683                 if (map->sm_fnv != fnv)
 684                         continue;
 685                 if (strcmp(map->sm_path, path) == 0) {
 686 #ifdef MAC
 687                         error = mac_posixshm_check_unlink(ucred, map->sm_shmfd);
 688                         if (error)
 689                                 return (error);
 690 #endif
 691                         error = shm_access(map->sm_shmfd, ucred,
 692                             FREAD | FWRITE);
 693                         if (error)
 694                                 return (error);
 695                         map->sm_shmfd->shm_path = NULL;
 696                         LIST_REMOVE(map, sm_link);
 697                         shm_drop(map->sm_shmfd);
 698                         free(map->sm_path, M_SHMFD);
 699                         free(map, M_SHMFD);
 700                         return (0);
 701                 }
 702         }
 703
 704         return (ENOENT);
 705 }
 706
 707 /* System calls. */
 708 int
 709 sys_shm_open(struct thread *td, struct shm_open_args *uap)
 710 {
 711         struct filedesc *fdp;
 712         struct shmfd *shmfd;
 713         struct file *fp;
 714         char *path;
 715         const char *pr_path;
 716         size_t pr_pathlen;
 717         Fnv32_t fnv;
 718         mode_t cmode;
 719         int fd, error;
 720
 721 #ifdef CAPABILITY_MODE
 722         /*
 723          * shm_open(2) is only allowed for anonymous objects.
 724          */
 725         if (IN_CAPABILITY_MODE(td) && (uap->path != SHM_ANON))
 726                 return (ECAPMODE);
 727 #endif
 728
 729         if ((uap->flags & O_ACCMODE) != O_RDONLY &&
 730             (uap->flags & O_ACCMODE) != O_RDWR)
 731                 return (EINVAL);
 732
 733         if ((uap->flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0)
 734                 return (EINVAL);
 735
 736         fdp = td->td_proc->p_fd;
 737         cmode = (uap->mode & ~fdp->fd_cmask) & ACCESSPERMS;
 738
 739         error = falloc(td, &fp, &fd, O_CLOEXEC);
 740         if (error)
 741                 return (error);
 742
 743         /* A SHM_ANON path pointer creates an anonymous object. */
 744         if (uap->path == SHM_ANON) {
 745                 /* A read-only anonymous object is pointless. */
 746                 if ((uap->flags & O_ACCMODE) == O_RDONLY) {
 747                         fdclose(fdp, fp, fd, td);
 748                         fdrop(fp, td);
 749                         return (EINVAL);
 750                 }
 751                 shmfd = shm_alloc(td->td_ucred, cmode);
 752         } else {
 753                 path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK);
 754                 pr_path = td->td_ucred->cr_prison->pr_path;
 755
 756                 /* Construct a full pathname for jailed callers. */
 757                 pr_pathlen = strcmp(pr_path, "/") == 0 ? 0
 758                     : strlcpy(path, pr_path, MAXPATHLEN);
 759                 error = copyinstr(uap->path, path + pr_pathlen,
 760                     MAXPATHLEN - pr_pathlen, NULL);
 761 #ifdef KTRACE
 762                 if (error == 0 && KTRPOINT(curthread, KTR_NAMEI))
 763                         ktrnamei(path);
 764 #endif
 765                 /* Require paths to start with a '/' character. */
 766                 if (error == 0 && path[pr_pathlen] != '/')
 767                         error = EINVAL;
 768                 if (error) {
 769                         fdclose(fdp, fp, fd, td);
 770                         fdrop(fp, td);
 771                         free(path, M_SHMFD);
 772                         return (error);
 773                 }
 774
 775                 fnv = fnv_32_str(path, FNV1_32_INIT);
 776                 sx_xlock(&shm_dict_lock);
 777                 shmfd = shm_lookup(path, fnv);
 778                 if (shmfd == NULL) {
 779                         /* Object does not yet exist, create it if requested. */
 780                         if (uap->flags & O_CREAT) {
 781 #ifdef MAC
 782                                 error = mac_posixshm_check_create(td->td_ucred,
 783                                     path);
 784                                 if (error == 0) {
 785 #endif
 786                                         shmfd = shm_alloc(td->td_ucred, cmode);
 787                                         shm_insert(path, fnv, shmfd);
 788 #ifdef MAC
 789                                 }
 790 #endif
 791                         } else {
 792                                 free(path, M_SHMFD);
 793                                 error = ENOENT;
 794                         }
 795                 } else {
 796                         /*
 797                          * Object already exists, obtain a new
 798                          * reference if requested and permitted.
 799                          */
 800                         free(path, M_SHMFD);
 801                         if ((uap->flags & (O_CREAT | O_EXCL)) ==
 802                             (O_CREAT | O_EXCL))
 803                                 error = EEXIST;
 804                         else {
 805 #ifdef MAC
 806                                 error = mac_posixshm_check_open(td->td_ucred,
 807                                     shmfd, FFLAGS(uap->flags & O_ACCMODE));
 808                                 if (error == 0)
 809 #endif
 810                                 error = shm_access(shmfd, td->td_ucred,
 811                                     FFLAGS(uap->flags & O_ACCMODE));
 812                         }
 813
 814                         /*
 815                          * Truncate the file back to zero length if
 816                          * O_TRUNC was specified and the object was
 817                          * opened with read/write.
 818                          */
 819                         if (error == 0 &&
 820                             (uap->flags & (O_ACCMODE | O_TRUNC)) ==
 821                             (O_RDWR | O_TRUNC)) {
 822 #ifdef MAC
 823                                 error = mac_posixshm_check_truncate(
 824                                         td->td_ucred, fp->f_cred, shmfd);
 825                                 if (error == 0)
 826 #endif
 827                                         shm_dotruncate(shmfd, 0);
 828                         }
 829                         if (error == 0)
 830                                 shm_hold(shmfd);
 831                 }
 832                 sx_xunlock(&shm_dict_lock);
 833
 834                 if (error) {
 835                         fdclose(fdp, fp, fd, td);
 836                         fdrop(fp, td);
 837                         return (error);
 838                 }
 839         }
 840
 841         finit(fp, FFLAGS(uap->flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops);
 842
 843         td->td_retval[0] = fd;
 844         fdrop(fp, td);
 845
 846         return (0);
 847 }
 848
 849 int
 850 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap)
 851 {
 852         char *path;
 853         const char *pr_path;
 854         size_t pr_pathlen;
 855         Fnv32_t fnv;
 856         int error;
 857
 858         path = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
 859         pr_path = td->td_ucred->cr_prison->pr_path;
 860         pr_pathlen = strcmp(pr_path, "/") == 0 ? 0
 861             : strlcpy(path, pr_path, MAXPATHLEN);
 862         error = copyinstr(uap->path, path + pr_pathlen, MAXPATHLEN - pr_pathlen,
 863             NULL);
 864         if (error) {
 865                 free(path, M_TEMP);
 866                 return (error);
 867         }
 868 #ifdef KTRACE
 869         if (KTRPOINT(curthread, KTR_NAMEI))
 870                 ktrnamei(path);
 871 #endif
 872         fnv = fnv_32_str(path, FNV1_32_INIT);
 873         sx_xlock(&shm_dict_lock);
 874         error = shm_remove(path, fnv, td->td_ucred);
 875         sx_xunlock(&shm_dict_lock);
 876         free(path, M_TEMP);
 877
 878         return (error);
 879 }
 880
 881 /*
 882  * mmap() helper to validate mmap() requests against shm object state
 883  * and give mmap() the vm_object to use for the mapping.
 884  */
 885 int
 886 shm_mmap(struct shmfd *shmfd, vm_size_t objsize, vm_ooffset_t foff,
 887     vm_object_t *obj)
 888 {
 889
 890         /*
 891          * XXXRW: This validation is probably insufficient, and subject to
 892          * sign errors.  It should be fixed.
 893          */
 894         if (foff >= shmfd->shm_size ||
 895             foff + objsize > round_page(shmfd->shm_size))
 896                 return (EINVAL);
 897
 898         mtx_lock(&shm_timestamp_lock);
 899         vfs_timestamp(&shmfd->shm_atime);
 900         mtx_unlock(&shm_timestamp_lock);
 901         vm_object_reference(shmfd->shm_object);
 902         *obj = shmfd->shm_object;
 903         return (0);
 904 }
 905
 906 static int
 907 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
 908     struct thread *td)
 909 {
 910         struct shmfd *shmfd;
 911         int error;
 912
 913         error = 0;
 914         shmfd = fp->f_data;
 915         mtx_lock(&shm_timestamp_lock);
 916         /*
 917          * SUSv4 says that x bits of permission need not be affected.
 918          * Be consistent with our shm_open there.
 919          */
 920 #ifdef MAC
 921         error = mac_posixshm_check_setmode(active_cred, shmfd, mode);
 922         if (error != 0)
 923                 goto out;
 924 #endif
 925         error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid,
 926             shmfd->shm_gid, VADMIN, active_cred, NULL);
 927         if (error != 0)
 928                 goto out;
 929         shmfd->shm_mode = mode & ACCESSPERMS;
 930 out:
 931         mtx_unlock(&shm_timestamp_lock);
 932         return (error);
 933 }
 934
 935 static int
 936 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
 937     struct thread *td)
 938 {
 939         struct shmfd *shmfd;
 940         int error;
 941
 942         error = 0;
 943         shmfd = fp->f_data;
 944         mtx_lock(&shm_timestamp_lock);
 945 #ifdef MAC
 946         error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid);
 947         if (error != 0)
 948                 goto out;
 949 #endif
 950         if (uid == (uid_t)-1)
 951                 uid = shmfd->shm_uid;
 952         if (gid == (gid_t)-1)
 953                  gid = shmfd->shm_gid;
 954         if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) ||
 955             (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) &&
 956             (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN, 0)))
 957                 goto out;
 958         shmfd->shm_uid = uid;
 959         shmfd->shm_gid = gid;
 960 out:
 961         mtx_unlock(&shm_timestamp_lock);
 962         return (error);
 963 }
 964
 965 /*
 966  * Helper routines to allow the backing object of a shared memory file
 967  * descriptor to be mapped in the kernel.
 968  */
 969 int
 970 shm_map(struct file *fp, size_t size, off_t offset, void **memp)
 971 {
 972         struct shmfd *shmfd;
 973         vm_offset_t kva, ofs;
 974         vm_object_t obj;
 975         int rv;
 976
 977         if (fp->f_type != DTYPE_SHM)
 978                 return (EINVAL);
 979         shmfd = fp->f_data;
 980         obj = shmfd->shm_object;
 981         VM_OBJECT_WLOCK(obj);
 982         /*
 983          * XXXRW: This validation is probably insufficient, and subject to
 984          * sign errors.  It should be fixed.
 985          */
 986         if (offset >= shmfd->shm_size ||
 987             offset + size > round_page(shmfd->shm_size)) {
 988                 VM_OBJECT_WUNLOCK(obj);
 989                 return (EINVAL);
 990         }
 991
 992         shmfd->shm_kmappings++;
 993         vm_object_reference_locked(obj);
 994         VM_OBJECT_WUNLOCK(obj);
 995
 996         /* Map the object into the kernel_map and wire it. */
 997         kva = vm_map_min(kernel_map);
 998         ofs = offset & PAGE_MASK;
 999         offset = trunc_page(offset);
1000         size = round_page(size + ofs);
1001         rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0,
1002             VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE,
1003             VM_PROT_READ | VM_PROT_WRITE, 0);
1004         if (rv == KERN_SUCCESS) {
1005                 rv = vm_map_wire(kernel_map, kva, kva + size,
1006                     VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
1007                 if (rv == KERN_SUCCESS) {
1008                         *memp = (void *)(kva + ofs);
1009                         return (0);
1010                 }
1011                 vm_map_remove(kernel_map, kva, kva + size);
1012         } else
1013                 vm_object_deallocate(obj);
1014
1015         /* On failure, drop our mapping reference. */
1016         VM_OBJECT_WLOCK(obj);
1017         shmfd->shm_kmappings--;
1018         VM_OBJECT_WUNLOCK(obj);
1019
1020         return (vm_mmap_to_errno(rv));
1021 }
1022
1023 /*
1024  * We require the caller to unmap the entire entry.  This allows us to
1025  * safely decrement shm_kmappings when a mapping is removed.
1026  */
1027 int
1028 shm_unmap(struct file *fp, void *mem, size_t size)
1029 {
1030         struct shmfd *shmfd;
1031         vm_map_entry_t entry;
1032         vm_offset_t kva, ofs;
1033         vm_object_t obj;
1034         vm_pindex_t pindex;
1035         vm_prot_t prot;
1036         boolean_t wired;
1037         vm_map_t map;
1038         int rv;
1039
1040         if (fp->f_type != DTYPE_SHM)
1041                 return (EINVAL);
1042         shmfd = fp->f_data;
1043         kva = (vm_offset_t)mem;
1044         ofs = kva & PAGE_MASK;
1045         kva = trunc_page(kva);
1046         size = round_page(size + ofs);
1047         map = kernel_map;
1048         rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry,
1049             &obj, &pindex, &prot, &wired);
1050         if (rv != KERN_SUCCESS)
1051                 return (EINVAL);
1052         if (entry->start != kva || entry->end != kva + size) {
1053                 vm_map_lookup_done(map, entry);
1054                 return (EINVAL);
1055         }
1056         vm_map_lookup_done(map, entry);
1057         if (obj != shmfd->shm_object)
1058                 return (EINVAL);
1059         vm_map_remove(map, kva, kva + size);
1060         VM_OBJECT_WLOCK(obj);
1061         KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped"));
1062         shmfd->shm_kmappings--;
1063         VM_OBJECT_WUNLOCK(obj);
1064         return (0);
1065 }
1066
1067 void
1068 shm_path(struct shmfd *shmfd, char *path, size_t size)
1069 {
1070         const char *shm_path, *pr_path;
1071         size_t pr_pathlen;
1072
1073         if (shmfd->shm_path == NULL)
1074                 return;
1075         sx_slock(&shm_dict_lock);
1076         shm_path = shmfd->shm_path;
1077         if (shm_path != NULL) {
1078                 pr_path = curthread->td_ucred->cr_prison->pr_path;
1079                 if (strcmp(pr_path, "/") != 0) {
1080                         /* Return the jail-rooted pathname. */
1081                         pr_pathlen = strlen(pr_path);
1082                         if (strncmp(shm_path, pr_path, pr_pathlen) == 0 &&
1083                             shm_path[pr_pathlen] == '/')
1084                                 shm_path += pr_pathlen;
1085                 }
1086                 strlcpy(path, shm_path, size);
1087         }
1088         sx_sunlock(&shm_dict_lock);
1089 }