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 #ifdef MAC
 298         error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd);
 299         if (error)
 300                 return (error);
 301 #endif
 302         foffset_lock_uio(fp, uio, flags);
 303         rl_cookie = rangelock_rlock(&shmfd->shm_rl, uio->uio_offset,
 304             uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
 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         KASSERT(length >= 0, ("shm_dotruncate: length < 0"));
 443         object = shmfd->shm_object;
 444         VM_OBJECT_WLOCK(object);
 445         if (length == shmfd->shm_size) {
 446                 VM_OBJECT_WUNLOCK(object);
 447                 return (0);
 448         }
 449         nobjsize = OFF_TO_IDX(length + PAGE_MASK);
 450
 451         /* Are we shrinking?  If so, trim the end. */
 452         if (length < shmfd->shm_size) {
 453                 /*
 454                  * Disallow any requests to shrink the size if this
 455                  * object is mapped into the kernel.
 456                  */
 457                 if (shmfd->shm_kmappings > 0) {
 458                         VM_OBJECT_WUNLOCK(object);
 459                         return (EBUSY);
 460                 }
 461
 462                 /*
 463                  * Zero the truncated part of the last page.
 464                  */
 465                 base = length & PAGE_MASK;
 466                 if (base != 0) {
 467                         idx = OFF_TO_IDX(length);
 468 retry:
 469                         m = vm_page_lookup(object, idx);
 470                         if (m != NULL) {
 471                                 if (vm_page_sleep_if_busy(m, "shmtrc"))
 472                                         goto retry;
 473                         } else if (vm_pager_has_page(object, idx, NULL, NULL)) {
 474                                 m = vm_page_alloc(object, idx, VM_ALLOC_NORMAL);
 475                                 if (m == NULL) {
 476                                         VM_OBJECT_WUNLOCK(object);
 477                                         VM_WAIT;
 478                                         VM_OBJECT_WLOCK(object);
 479                                         goto retry;
 480                                 } else if (m->valid != VM_PAGE_BITS_ALL) {
 481                                         ma[0] = m;
 482                                         rv = vm_pager_get_pages(object, ma, 1,
 483                                             0);
 484                                         m = vm_page_lookup(object, idx);
 485                                 } else
 486                                         /* A cached page was reactivated. */
 487                                         rv = VM_PAGER_OK;
 488                                 vm_page_lock(m);
 489                                 if (rv == VM_PAGER_OK) {
 490                                         vm_page_deactivate(m);
 491                                         vm_page_unlock(m);
 492                                         vm_page_xunbusy(m);
 493                                 } else {
 494                                         vm_page_free(m);
 495                                         vm_page_unlock(m);
 496                                         VM_OBJECT_WUNLOCK(object);
 497                                         return (EIO);
 498                                 }
 499                         }
 500                         if (m != NULL) {
 501                                 pmap_zero_page_area(m, base, PAGE_SIZE - base);
 502                                 KASSERT(m->valid == VM_PAGE_BITS_ALL,
 503                                     ("shm_dotruncate: page %p is invalid", m));
 504                                 vm_page_dirty(m);
 505                                 vm_pager_page_unswapped(m);
 506                         }
 507                 }
 508                 delta = IDX_TO_OFF(object->size - nobjsize);
 509
 510                 /* Toss in memory pages. */
 511                 if (nobjsize < object->size)
 512                         vm_object_page_remove(object, nobjsize, object->size,
 513                             0);
 514
 515                 /* Toss pages from swap. */
 516                 if (object->type == OBJT_SWAP)
 517                         swap_pager_freespace(object, nobjsize, delta);
 518
 519                 /* Free the swap accounted for shm */
 520                 swap_release_by_cred(delta, object->cred);
 521                 object->charge -= delta;
 522         } else {
 523                 /* Try to reserve additional swap space. */
 524                 delta = IDX_TO_OFF(nobjsize - object->size);
 525                 if (!swap_reserve_by_cred(delta, object->cred)) {
 526                         VM_OBJECT_WUNLOCK(object);
 527                         return (ENOMEM);
 528                 }
 529                 object->charge += delta;
 530         }
 531         shmfd->shm_size = length;
 532         mtx_lock(&shm_timestamp_lock);
 533         vfs_timestamp(&shmfd->shm_ctime);
 534         shmfd->shm_mtime = shmfd->shm_ctime;
 535         mtx_unlock(&shm_timestamp_lock);
 536         object->size = nobjsize;
 537         VM_OBJECT_WUNLOCK(object);
 538         return (0);
 539 }
 540
 541 /*
 542  * shmfd object management including creation and reference counting
 543  * routines.
 544  */
 545 static struct shmfd *
 546 shm_alloc(struct ucred *ucred, mode_t mode)
 547 {
 548         struct shmfd *shmfd;
 549         int ino;
 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         ino = alloc_unr(shm_ino_unr);
 568         if (ino == -1)
 569                 shmfd->shm_ino = 0;
 570         else
 571                 shmfd->shm_ino = ino;
 572         refcount_init(&shmfd->shm_refs, 1);
 573         mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF);
 574         rangelock_init(&shmfd->shm_rl);
 575 #ifdef MAC
 576         mac_posixshm_init(shmfd);
 577         mac_posixshm_create(ucred, shmfd);
 578 #endif
 579
 580         return (shmfd);
 581 }
 582
 583 static struct shmfd *
 584 shm_hold(struct shmfd *shmfd)
 585 {
 586
 587         refcount_acquire(&shmfd->shm_refs);
 588         return (shmfd);
 589 }
 590
 591 static void
 592 shm_drop(struct shmfd *shmfd)
 593 {
 594
 595         if (refcount_release(&shmfd->shm_refs)) {
 596 #ifdef MAC
 597                 mac_posixshm_destroy(shmfd);
 598 #endif
 599                 rangelock_destroy(&shmfd->shm_rl);
 600                 mtx_destroy(&shmfd->shm_mtx);
 601                 vm_object_deallocate(shmfd->shm_object);
 602                 if (shmfd->shm_ino != 0)
 603                         free_unr(shm_ino_unr, shmfd->shm_ino);
 604                 free(shmfd, M_SHMFD);
 605         }
 606 }
 607
 608 /*
 609  * Determine if the credentials have sufficient permissions for a
 610  * specified combination of FREAD and FWRITE.
 611  */
 612 static int
 613 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags)
 614 {
 615         accmode_t accmode;
 616         int error;
 617
 618         accmode = 0;
 619         if (flags & FREAD)
 620                 accmode |= VREAD;
 621         if (flags & FWRITE)
 622                 accmode |= VWRITE;
 623         mtx_lock(&shm_timestamp_lock);
 624         error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
 625             accmode, ucred, NULL);
 626         mtx_unlock(&shm_timestamp_lock);
 627         return (error);
 628 }
 629
 630 /*
 631  * Dictionary management.  We maintain an in-kernel dictionary to map
 632  * paths to shmfd objects.  We use the FNV hash on the path to store
 633  * the mappings in a hash table.
 634  */
 635 static void
 636 shm_init(void *arg)
 637 {
 638
 639         mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF);
 640         sx_init(&shm_dict_lock, "shm dictionary");
 641         shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash);
 642         shm_ino_unr = new_unrhdr(1, INT32_MAX, NULL);
 643         KASSERT(shm_ino_unr != NULL, ("shm fake inodes not initialized"));
 644         shm_dev_ino = devfs_alloc_cdp_inode();
 645         KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized"));
 646 }
 647 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL);
 648
 649 static struct shmfd *
 650 shm_lookup(char *path, Fnv32_t fnv)
 651 {
 652         struct shm_mapping *map;
 653
 654         LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
 655                 if (map->sm_fnv != fnv)
 656                         continue;
 657                 if (strcmp(map->sm_path, path) == 0)
 658                         return (map->sm_shmfd);
 659         }
 660
 661         return (NULL);
 662 }
 663
 664 static void
 665 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd)
 666 {
 667         struct shm_mapping *map;
 668
 669         map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK);
 670         map->sm_path = path;
 671         map->sm_fnv = fnv;
 672         map->sm_shmfd = shm_hold(shmfd);
 673         shmfd->shm_path = path;
 674         LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link);
 675 }
 676
 677 static int
 678 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred)
 679 {
 680         struct shm_mapping *map;
 681         int error;
 682
 683         LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
 684                 if (map->sm_fnv != fnv)
 685                         continue;
 686                 if (strcmp(map->sm_path, path) == 0) {
 687 #ifdef MAC
 688                         error = mac_posixshm_check_unlink(ucred, map->sm_shmfd);
 689                         if (error)
 690                                 return (error);
 691 #endif
 692                         error = shm_access(map->sm_shmfd, ucred,
 693                             FREAD | FWRITE);
 694                         if (error)
 695                                 return (error);
 696                         map->sm_shmfd->shm_path = NULL;
 697                         LIST_REMOVE(map, sm_link);
 698                         shm_drop(map->sm_shmfd);
 699                         free(map->sm_path, M_SHMFD);
 700                         free(map, M_SHMFD);
 701                         return (0);
 702                 }
 703         }
 704
 705         return (ENOENT);
 706 }
 707
 708 /* System calls. */
 709 int
 710 sys_shm_open(struct thread *td, struct shm_open_args *uap)
 711 {
 712         struct filedesc *fdp;
 713         struct shmfd *shmfd;
 714         struct file *fp;
 715         char *path;
 716         const char *pr_path;
 717         size_t pr_pathlen;
 718         Fnv32_t fnv;
 719         mode_t cmode;
 720         int fd, error;
 721
 722 #ifdef CAPABILITY_MODE
 723         /*
 724          * shm_open(2) is only allowed for anonymous objects.
 725          */
 726         if (IN_CAPABILITY_MODE(td) && (uap->path != SHM_ANON))
 727                 return (ECAPMODE);
 728 #endif
 729
 730         if ((uap->flags & O_ACCMODE) != O_RDONLY &&
 731             (uap->flags & O_ACCMODE) != O_RDWR)
 732                 return (EINVAL);
 733
 734         if ((uap->flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0)
 735                 return (EINVAL);
 736
 737         fdp = td->td_proc->p_fd;
 738         cmode = (uap->mode & ~fdp->fd_cmask) & ACCESSPERMS;
 739
 740         error = falloc(td, &fp, &fd, O_CLOEXEC);
 741         if (error)
 742                 return (error);
 743
 744         /* A SHM_ANON path pointer creates an anonymous object. */
 745         if (uap->path == SHM_ANON) {
 746                 /* A read-only anonymous object is pointless. */
 747                 if ((uap->flags & O_ACCMODE) == O_RDONLY) {
 748                         fdclose(td, fp, fd);
 749                         fdrop(fp, td);
 750                         return (EINVAL);
 751                 }
 752                 shmfd = shm_alloc(td->td_ucred, cmode);
 753         } else {
 754                 path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK);
 755                 pr_path = td->td_ucred->cr_prison->pr_path;
 756
 757                 /* Construct a full pathname for jailed callers. */
 758                 pr_pathlen = strcmp(pr_path, "/") == 0 ? 0
 759                     : strlcpy(path, pr_path, MAXPATHLEN);
 760                 error = copyinstr(uap->path, path + pr_pathlen,
 761                     MAXPATHLEN - pr_pathlen, NULL);
 762 #ifdef KTRACE
 763                 if (error == 0 && KTRPOINT(curthread, KTR_NAMEI))
 764                         ktrnamei(path);
 765 #endif
 766                 /* Require paths to start with a '/' character. */
 767                 if (error == 0 && path[pr_pathlen] != '/')
 768                         error = EINVAL;
 769                 if (error) {
 770                         fdclose(td, fp, fd);
 771                         fdrop(fp, td);
 772                         free(path, M_SHMFD);
 773                         return (error);
 774                 }
 775
 776                 fnv = fnv_32_str(path, FNV1_32_INIT);
 777                 sx_xlock(&shm_dict_lock);
 778                 shmfd = shm_lookup(path, fnv);
 779                 if (shmfd == NULL) {
 780                         /* Object does not yet exist, create it if requested. */
 781                         if (uap->flags & O_CREAT) {
 782 #ifdef MAC
 783                                 error = mac_posixshm_check_create(td->td_ucred,
 784                                     path);
 785                                 if (error == 0) {
 786 #endif
 787                                         shmfd = shm_alloc(td->td_ucred, cmode);
 788                                         shm_insert(path, fnv, shmfd);
 789 #ifdef MAC
 790                                 }
 791 #endif
 792                         } else {
 793                                 free(path, M_SHMFD);
 794                                 error = ENOENT;
 795                         }
 796                 } else {
 797                         /*
 798                          * Object already exists, obtain a new
 799                          * reference if requested and permitted.
 800                          */
 801                         free(path, M_SHMFD);
 802                         if ((uap->flags & (O_CREAT | O_EXCL)) ==
 803                             (O_CREAT | O_EXCL))
 804                                 error = EEXIST;
 805                         else {
 806 #ifdef MAC
 807                                 error = mac_posixshm_check_open(td->td_ucred,
 808                                     shmfd, FFLAGS(uap->flags & O_ACCMODE));
 809                                 if (error == 0)
 810 #endif
 811                                 error = shm_access(shmfd, td->td_ucred,
 812                                     FFLAGS(uap->flags & O_ACCMODE));
 813                         }
 814
 815                         /*
 816                          * Truncate the file back to zero length if
 817                          * O_TRUNC was specified and the object was
 818                          * opened with read/write.
 819                          */
 820                         if (error == 0 &&
 821                             (uap->flags & (O_ACCMODE | O_TRUNC)) ==
 822                             (O_RDWR | O_TRUNC)) {
 823 #ifdef MAC
 824                                 error = mac_posixshm_check_truncate(
 825                                         td->td_ucred, fp->f_cred, shmfd);
 826                                 if (error == 0)
 827 #endif
 828                                         shm_dotruncate(shmfd, 0);
 829                         }
 830                         if (error == 0)
 831                                 shm_hold(shmfd);
 832                 }
 833                 sx_xunlock(&shm_dict_lock);
 834
 835                 if (error) {
 836                         fdclose(td, fp, fd);
 837                         fdrop(fp, td);
 838                         return (error);
 839                 }
 840         }
 841
 842         finit(fp, FFLAGS(uap->flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops);
 843
 844         td->td_retval[0] = fd;
 845         fdrop(fp, td);
 846
 847         return (0);
 848 }
 849
 850 int
 851 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap)
 852 {
 853         char *path;
 854         const char *pr_path;
 855         size_t pr_pathlen;
 856         Fnv32_t fnv;
 857         int error;
 858
 859         path = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
 860         pr_path = td->td_ucred->cr_prison->pr_path;
 861         pr_pathlen = strcmp(pr_path, "/") == 0 ? 0
 862             : strlcpy(path, pr_path, MAXPATHLEN);
 863         error = copyinstr(uap->path, path + pr_pathlen, MAXPATHLEN - pr_pathlen,
 864             NULL);
 865         if (error) {
 866                 free(path, M_TEMP);
 867                 return (error);
 868         }
 869 #ifdef KTRACE
 870         if (KTRPOINT(curthread, KTR_NAMEI))
 871                 ktrnamei(path);
 872 #endif
 873         fnv = fnv_32_str(path, FNV1_32_INIT);
 874         sx_xlock(&shm_dict_lock);
 875         error = shm_remove(path, fnv, td->td_ucred);
 876         sx_xunlock(&shm_dict_lock);
 877         free(path, M_TEMP);
 878
 879         return (error);
 880 }
 881
 882 /*
 883  * mmap() helper to validate mmap() requests against shm object state
 884  * and give mmap() the vm_object to use for the mapping.
 885  */
 886 int
 887 shm_mmap(struct shmfd *shmfd, vm_size_t objsize, vm_ooffset_t foff,
 888     vm_object_t *obj)
 889 {
 890
 891         /*
 892          * XXXRW: This validation is probably insufficient, and subject to
 893          * sign errors.  It should be fixed.
 894          */
 895         if (foff >= shmfd->shm_size ||
 896             foff + objsize > round_page(shmfd->shm_size))
 897                 return (EINVAL);
 898
 899         mtx_lock(&shm_timestamp_lock);
 900         vfs_timestamp(&shmfd->shm_atime);
 901         mtx_unlock(&shm_timestamp_lock);
 902         vm_object_reference(shmfd->shm_object);
 903         *obj = shmfd->shm_object;
 904         return (0);
 905 }
 906
 907 static int
 908 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
 909     struct thread *td)
 910 {
 911         struct shmfd *shmfd;
 912         int error;
 913
 914         error = 0;
 915         shmfd = fp->f_data;
 916         mtx_lock(&shm_timestamp_lock);
 917         /*
 918          * SUSv4 says that x bits of permission need not be affected.
 919          * Be consistent with our shm_open there.
 920          */
 921 #ifdef MAC
 922         error = mac_posixshm_check_setmode(active_cred, shmfd, mode);
 923         if (error != 0)
 924                 goto out;
 925 #endif
 926         error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid,
 927             shmfd->shm_gid, VADMIN, active_cred, NULL);
 928         if (error != 0)
 929                 goto out;
 930         shmfd->shm_mode = mode & ACCESSPERMS;
 931 out:
 932         mtx_unlock(&shm_timestamp_lock);
 933         return (error);
 934 }
 935
 936 static int
 937 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
 938     struct thread *td)
 939 {
 940         struct shmfd *shmfd;
 941         int error;
 942
 943         error = 0;
 944         shmfd = fp->f_data;
 945         mtx_lock(&shm_timestamp_lock);
 946 #ifdef MAC
 947         error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid);
 948         if (error != 0)
 949                 goto out;
 950 #endif
 951         if (uid == (uid_t)-1)
 952                 uid = shmfd->shm_uid;
 953         if (gid == (gid_t)-1)
 954                  gid = shmfd->shm_gid;
 955         if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) ||
 956             (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) &&
 957             (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN, 0)))
 958                 goto out;
 959         shmfd->shm_uid = uid;
 960         shmfd->shm_gid = gid;
 961 out:
 962         mtx_unlock(&shm_timestamp_lock);
 963         return (error);
 964 }
 965
 966 /*
 967  * Helper routines to allow the backing object of a shared memory file
 968  * descriptor to be mapped in the kernel.
 969  */
 970 int
 971 shm_map(struct file *fp, size_t size, off_t offset, void **memp)
 972 {
 973         struct shmfd *shmfd;
 974         vm_offset_t kva, ofs;
 975         vm_object_t obj;
 976         int rv;
 977
 978         if (fp->f_type != DTYPE_SHM)
 979                 return (EINVAL);
 980         shmfd = fp->f_data;
 981         obj = shmfd->shm_object;
 982         VM_OBJECT_WLOCK(obj);
 983         /*
 984          * XXXRW: This validation is probably insufficient, and subject to
 985          * sign errors.  It should be fixed.
 986          */
 987         if (offset >= shmfd->shm_size ||
 988             offset + size > round_page(shmfd->shm_size)) {
 989                 VM_OBJECT_WUNLOCK(obj);
 990                 return (EINVAL);
 991         }
 992
 993         shmfd->shm_kmappings++;
 994         vm_object_reference_locked(obj);
 995         VM_OBJECT_WUNLOCK(obj);
 996
 997         /* Map the object into the kernel_map and wire it. */
 998         kva = vm_map_min(kernel_map);
 999         ofs = offset & PAGE_MASK;
1000         offset = trunc_page(offset);
1001         size = round_page(size + ofs);
1002         rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0,
1003             VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE,
1004             VM_PROT_READ | VM_PROT_WRITE, 0);
1005         if (rv == KERN_SUCCESS) {
1006                 rv = vm_map_wire(kernel_map, kva, kva + size,
1007                     VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
1008                 if (rv == KERN_SUCCESS) {
1009                         *memp = (void *)(kva + ofs);
1010                         return (0);
1011                 }
1012                 vm_map_remove(kernel_map, kva, kva + size);
1013         } else
1014                 vm_object_deallocate(obj);
1015
1016         /* On failure, drop our mapping reference. */
1017         VM_OBJECT_WLOCK(obj);
1018         shmfd->shm_kmappings--;
1019         VM_OBJECT_WUNLOCK(obj);
1020
1021         return (vm_mmap_to_errno(rv));
1022 }
1023
1024 /*
1025  * We require the caller to unmap the entire entry.  This allows us to
1026  * safely decrement shm_kmappings when a mapping is removed.
1027  */
1028 int
1029 shm_unmap(struct file *fp, void *mem, size_t size)
1030 {
1031         struct shmfd *shmfd;
1032         vm_map_entry_t entry;
1033         vm_offset_t kva, ofs;
1034         vm_object_t obj;
1035         vm_pindex_t pindex;
1036         vm_prot_t prot;
1037         boolean_t wired;
1038         vm_map_t map;
1039         int rv;
1040
1041         if (fp->f_type != DTYPE_SHM)
1042                 return (EINVAL);
1043         shmfd = fp->f_data;
1044         kva = (vm_offset_t)mem;
1045         ofs = kva & PAGE_MASK;
1046         kva = trunc_page(kva);
1047         size = round_page(size + ofs);
1048         map = kernel_map;
1049         rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry,
1050             &obj, &pindex, &prot, &wired);
1051         if (rv != KERN_SUCCESS)
1052                 return (EINVAL);
1053         if (entry->start != kva || entry->end != kva + size) {
1054                 vm_map_lookup_done(map, entry);
1055                 return (EINVAL);
1056         }
1057         vm_map_lookup_done(map, entry);
1058         if (obj != shmfd->shm_object)
1059                 return (EINVAL);
1060         vm_map_remove(map, kva, kva + size);
1061         VM_OBJECT_WLOCK(obj);
1062         KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped"));
1063         shmfd->shm_kmappings--;
1064         VM_OBJECT_WUNLOCK(obj);
1065         return (0);
1066 }
1067
1068 void
1069 shm_path(struct shmfd *shmfd, char *path, size_t size)
1070 {
1071         const char *shm_path, *pr_path;
1072         size_t pr_pathlen;
1073
1074         if (shmfd->shm_path == NULL)
1075                 return;
1076         sx_slock(&shm_dict_lock);
1077         shm_path = shmfd->shm_path;
1078         if (shm_path != NULL) {
1079                 pr_path = curthread->td_ucred->cr_prison->pr_path;
1080                 if (strcmp(pr_path, "/") != 0) {
1081                         /* Return the jail-rooted pathname. */
1082                         pr_pathlen = strlen(pr_path);
1083                         if (strncmp(shm_path, pr_path, pr_pathlen) == 0 &&
1084                             shm_path[pr_pathlen] == '/')
1085                                 shm_path += pr_pathlen;
1086                 }
1087                 strlcpy(path, shm_path, size);
1088         }
1089         sx_sunlock(&shm_dict_lock);
1090 }