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/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 #include <sys/user.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_stat_t        shm_stat;
 125 static fo_close_t       shm_close;
 126 static fo_chmod_t       shm_chmod;
 127 static fo_chown_t       shm_chown;
 128 static fo_seek_t        shm_seek;
 129 static fo_fill_kinfo_t  shm_fill_kinfo;
 130
 131 /* File descriptor operations. */
 132 static struct fileops shm_ops = {
 133         .fo_read = shm_read,
 134         .fo_write = shm_write,
 135         .fo_truncate = shm_truncate,
 136         .fo_ioctl = invfo_ioctl,
 137         .fo_poll = invfo_poll,
 138         .fo_kqfilter = invfo_kqfilter,
 139         .fo_stat = shm_stat,
 140         .fo_close = shm_close,
 141         .fo_chmod = shm_chmod,
 142         .fo_chown = shm_chown,
 143         .fo_sendfile = vn_sendfile,
 144         .fo_seek = shm_seek,
 145         .fo_fill_kinfo = shm_fill_kinfo,
 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                         td->td_uretoff.tdu_off = 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_stat(struct file *fp, struct stat *sb, struct ucred *active_cred,
 359     struct thread *td)
 360 {
 361         struct shmfd *shmfd;
 362 #ifdef MAC
 363         int error;
 364 #endif
 365
 366         shmfd = fp->f_data;
 367
 368 #ifdef MAC
 369         error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd);
 370         if (error)
 371                 return (error);
 372 #endif
 373
 374         /*
 375          * Attempt to return sanish values for fstat() on a memory file
 376          * descriptor.
 377          */
 378         bzero(sb, sizeof(*sb));
 379         sb->st_blksize = PAGE_SIZE;
 380         sb->st_size = shmfd->shm_size;
 381         sb->st_blocks = (sb->st_size + sb->st_blksize - 1) / sb->st_blksize;
 382         mtx_lock(&shm_timestamp_lock);
 383         sb->st_atim = shmfd->shm_atime;
 384         sb->st_ctim = shmfd->shm_ctime;
 385         sb->st_mtim = shmfd->shm_mtime;
 386         sb->st_birthtim = shmfd->shm_birthtime;
 387         sb->st_mode = S_IFREG | shmfd->shm_mode;                /* XXX */
 388         sb->st_uid = shmfd->shm_uid;
 389         sb->st_gid = shmfd->shm_gid;
 390         mtx_unlock(&shm_timestamp_lock);
 391         sb->st_dev = shm_dev_ino;
 392         sb->st_ino = shmfd->shm_ino;
 393
 394         return (0);
 395 }
 396
 397 static int
 398 shm_close(struct file *fp, struct thread *td)
 399 {
 400         struct shmfd *shmfd;
 401
 402         shmfd = fp->f_data;
 403         fp->f_data = NULL;
 404         shm_drop(shmfd);
 405
 406         return (0);
 407 }
 408
 409 static int
 410 shm_dotruncate(struct shmfd *shmfd, off_t length)
 411 {
 412         vm_object_t object;
 413         vm_page_t m, ma[1];
 414         vm_pindex_t idx, nobjsize;
 415         vm_ooffset_t delta;
 416         int base, rv;
 417
 418         object = shmfd->shm_object;
 419         VM_OBJECT_WLOCK(object);
 420         if (length == shmfd->shm_size) {
 421                 VM_OBJECT_WUNLOCK(object);
 422                 return (0);
 423         }
 424         nobjsize = OFF_TO_IDX(length + PAGE_MASK);
 425
 426         /* Are we shrinking?  If so, trim the end. */
 427         if (length < shmfd->shm_size) {
 428                 /*
 429                  * Disallow any requests to shrink the size if this
 430                  * object is mapped into the kernel.
 431                  */
 432                 if (shmfd->shm_kmappings > 0) {
 433                         VM_OBJECT_WUNLOCK(object);
 434                         return (EBUSY);
 435                 }
 436
 437                 /*
 438                  * Zero the truncated part of the last page.
 439                  */
 440                 base = length & PAGE_MASK;
 441                 if (base != 0) {
 442                         idx = OFF_TO_IDX(length);
 443 retry:
 444                         m = vm_page_lookup(object, idx);
 445                         if (m != NULL) {
 446                                 if (vm_page_sleep_if_busy(m, "shmtrc"))
 447                                         goto retry;
 448                         } else if (vm_pager_has_page(object, idx, NULL, NULL)) {
 449                                 m = vm_page_alloc(object, idx, VM_ALLOC_NORMAL);
 450                                 if (m == NULL) {
 451                                         VM_OBJECT_WUNLOCK(object);
 452                                         VM_WAIT;
 453                                         VM_OBJECT_WLOCK(object);
 454                                         goto retry;
 455                                 } else if (m->valid != VM_PAGE_BITS_ALL) {
 456                                         ma[0] = m;
 457                                         rv = vm_pager_get_pages(object, ma, 1,
 458                                             0);
 459                                         m = vm_page_lookup(object, idx);
 460                                 } else
 461                                         /* A cached page was reactivated. */
 462                                         rv = VM_PAGER_OK;
 463                                 vm_page_lock(m);
 464                                 if (rv == VM_PAGER_OK) {
 465                                         vm_page_deactivate(m);
 466                                         vm_page_unlock(m);
 467                                         vm_page_xunbusy(m);
 468                                 } else {
 469                                         vm_page_free(m);
 470                                         vm_page_unlock(m);
 471                                         VM_OBJECT_WUNLOCK(object);
 472                                         return (EIO);
 473                                 }
 474                         }
 475                         if (m != NULL) {
 476                                 pmap_zero_page_area(m, base, PAGE_SIZE - base);
 477                                 KASSERT(m->valid == VM_PAGE_BITS_ALL,
 478                                     ("shm_dotruncate: page %p is invalid", m));
 479                                 vm_page_dirty(m);
 480                                 vm_pager_page_unswapped(m);
 481                         }
 482                 }
 483                 delta = ptoa(object->size - nobjsize);
 484
 485                 /* Toss in memory pages. */
 486                 if (nobjsize < object->size)
 487                         vm_object_page_remove(object, nobjsize, object->size,
 488                             0);
 489
 490                 /* Toss pages from swap. */
 491                 if (object->type == OBJT_SWAP)
 492                         swap_pager_freespace(object, nobjsize, delta);
 493
 494                 /* Free the swap accounted for shm */
 495                 swap_release_by_cred(delta, object->cred);
 496                 object->charge -= delta;
 497         } else {
 498                 /* Attempt to reserve the swap */
 499                 delta = ptoa(nobjsize - object->size);
 500                 if (!swap_reserve_by_cred(delta, object->cred)) {
 501                         VM_OBJECT_WUNLOCK(object);
 502                         return (ENOMEM);
 503                 }
 504                 object->charge += delta;
 505         }
 506         shmfd->shm_size = length;
 507         mtx_lock(&shm_timestamp_lock);
 508         vfs_timestamp(&shmfd->shm_ctime);
 509         shmfd->shm_mtime = shmfd->shm_ctime;
 510         mtx_unlock(&shm_timestamp_lock);
 511         object->size = nobjsize;
 512         VM_OBJECT_WUNLOCK(object);
 513         return (0);
 514 }
 515
 516 /*
 517  * shmfd object management including creation and reference counting
 518  * routines.
 519  */
 520 static struct shmfd *
 521 shm_alloc(struct ucred *ucred, mode_t mode)
 522 {
 523         struct shmfd *shmfd;
 524         int ino;
 525
 526         shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO);
 527         shmfd->shm_size = 0;
 528         shmfd->shm_uid = ucred->cr_uid;
 529         shmfd->shm_gid = ucred->cr_gid;
 530         shmfd->shm_mode = mode;
 531         shmfd->shm_object = vm_pager_allocate(OBJT_DEFAULT, NULL,
 532             shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred);
 533         KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate"));
 534         VM_OBJECT_WLOCK(shmfd->shm_object);
 535         vm_object_clear_flag(shmfd->shm_object, OBJ_ONEMAPPING);
 536         vm_object_set_flag(shmfd->shm_object, OBJ_NOSPLIT);
 537         VM_OBJECT_WUNLOCK(shmfd->shm_object);
 538         vfs_timestamp(&shmfd->shm_birthtime);
 539         shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime =
 540             shmfd->shm_birthtime;
 541         ino = alloc_unr(shm_ino_unr);
 542         if (ino == -1)
 543                 shmfd->shm_ino = 0;
 544         else
 545                 shmfd->shm_ino = ino;
 546         refcount_init(&shmfd->shm_refs, 1);
 547         mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF);
 548         rangelock_init(&shmfd->shm_rl);
 549 #ifdef MAC
 550         mac_posixshm_init(shmfd);
 551         mac_posixshm_create(ucred, shmfd);
 552 #endif
 553
 554         return (shmfd);
 555 }
 556
 557 static struct shmfd *
 558 shm_hold(struct shmfd *shmfd)
 559 {
 560
 561         refcount_acquire(&shmfd->shm_refs);
 562         return (shmfd);
 563 }
 564
 565 static void
 566 shm_drop(struct shmfd *shmfd)
 567 {
 568
 569         if (refcount_release(&shmfd->shm_refs)) {
 570 #ifdef MAC
 571                 mac_posixshm_destroy(shmfd);
 572 #endif
 573                 rangelock_destroy(&shmfd->shm_rl);
 574                 mtx_destroy(&shmfd->shm_mtx);
 575                 vm_object_deallocate(shmfd->shm_object);
 576                 if (shmfd->shm_ino != 0)
 577                         free_unr(shm_ino_unr, shmfd->shm_ino);
 578                 free(shmfd, M_SHMFD);
 579         }
 580 }
 581
 582 /*
 583  * Determine if the credentials have sufficient permissions for a
 584  * specified combination of FREAD and FWRITE.
 585  */
 586 static int
 587 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags)
 588 {
 589         accmode_t accmode;
 590         int error;
 591
 592         accmode = 0;
 593         if (flags & FREAD)
 594                 accmode |= VREAD;
 595         if (flags & FWRITE)
 596                 accmode |= VWRITE;
 597         mtx_lock(&shm_timestamp_lock);
 598         error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
 599             accmode, ucred, NULL);
 600         mtx_unlock(&shm_timestamp_lock);
 601         return (error);
 602 }
 603
 604 /*
 605  * Dictionary management.  We maintain an in-kernel dictionary to map
 606  * paths to shmfd objects.  We use the FNV hash on the path to store
 607  * the mappings in a hash table.
 608  */
 609 static void
 610 shm_init(void *arg)
 611 {
 612
 613         mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF);
 614         sx_init(&shm_dict_lock, "shm dictionary");
 615         shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash);
 616         shm_ino_unr = new_unrhdr(1, INT32_MAX, NULL);
 617         KASSERT(shm_ino_unr != NULL, ("shm fake inodes not initialized"));
 618         shm_dev_ino = devfs_alloc_cdp_inode();
 619         KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized"));
 620 }
 621 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL);
 622
 623 static struct shmfd *
 624 shm_lookup(char *path, Fnv32_t fnv)
 625 {
 626         struct shm_mapping *map;
 627
 628         LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
 629                 if (map->sm_fnv != fnv)
 630                         continue;
 631                 if (strcmp(map->sm_path, path) == 0)
 632                         return (map->sm_shmfd);
 633         }
 634
 635         return (NULL);
 636 }
 637
 638 static void
 639 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd)
 640 {
 641         struct shm_mapping *map;
 642
 643         map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK);
 644         map->sm_path = path;
 645         map->sm_fnv = fnv;
 646         map->sm_shmfd = shm_hold(shmfd);
 647         shmfd->shm_path = path;
 648         LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link);
 649 }
 650
 651 static int
 652 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred)
 653 {
 654         struct shm_mapping *map;
 655         int error;
 656
 657         LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
 658                 if (map->sm_fnv != fnv)
 659                         continue;
 660                 if (strcmp(map->sm_path, path) == 0) {
 661 #ifdef MAC
 662                         error = mac_posixshm_check_unlink(ucred, map->sm_shmfd);
 663                         if (error)
 664                                 return (error);
 665 #endif
 666                         error = shm_access(map->sm_shmfd, ucred,
 667                             FREAD | FWRITE);
 668                         if (error)
 669                                 return (error);
 670                         map->sm_shmfd->shm_path = NULL;
 671                         LIST_REMOVE(map, sm_link);
 672                         shm_drop(map->sm_shmfd);
 673                         free(map->sm_path, M_SHMFD);
 674                         free(map, M_SHMFD);
 675                         return (0);
 676                 }
 677         }
 678
 679         return (ENOENT);
 680 }
 681
 682 /* System calls. */
 683 int
 684 sys_shm_open(struct thread *td, struct shm_open_args *uap)
 685 {
 686         struct filedesc *fdp;
 687         struct shmfd *shmfd;
 688         struct file *fp;
 689         char *path;
 690         Fnv32_t fnv;
 691         mode_t cmode;
 692         int fd, error;
 693
 694 #ifdef CAPABILITY_MODE
 695         /*
 696          * shm_open(2) is only allowed for anonymous objects.
 697          */
 698         if (IN_CAPABILITY_MODE(td) && (uap->path != SHM_ANON))
 699                 return (ECAPMODE);
 700 #endif
 701
 702         if ((uap->flags & O_ACCMODE) != O_RDONLY &&
 703             (uap->flags & O_ACCMODE) != O_RDWR)
 704                 return (EINVAL);
 705
 706         if ((uap->flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0)
 707                 return (EINVAL);
 708
 709         fdp = td->td_proc->p_fd;
 710         cmode = (uap->mode & ~fdp->fd_cmask) & ACCESSPERMS;
 711
 712         error = falloc(td, &fp, &fd, O_CLOEXEC);
 713         if (error)
 714                 return (error);
 715
 716         /* A SHM_ANON path pointer creates an anonymous object. */
 717         if (uap->path == SHM_ANON) {
 718                 /* A read-only anonymous object is pointless. */
 719                 if ((uap->flags & O_ACCMODE) == O_RDONLY) {
 720                         fdclose(fdp, fp, fd, td);
 721                         fdrop(fp, td);
 722                         return (EINVAL);
 723                 }
 724                 shmfd = shm_alloc(td->td_ucred, cmode);
 725         } else {
 726                 path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK);
 727                 error = copyinstr(uap->path, path, MAXPATHLEN, NULL);
 728 #ifdef KTRACE
 729                 if (error == 0 && KTRPOINT(curthread, KTR_NAMEI))
 730                         ktrnamei(path);
 731 #endif
 732                 /* Require paths to start with a '/' character. */
 733                 if (error == 0 && path[0] != '/')
 734                         error = EINVAL;
 735                 if (error) {
 736                         fdclose(fdp, fp, fd, td);
 737                         fdrop(fp, td);
 738                         free(path, M_SHMFD);
 739                         return (error);
 740                 }
 741
 742                 fnv = fnv_32_str(path, FNV1_32_INIT);
 743                 sx_xlock(&shm_dict_lock);
 744                 shmfd = shm_lookup(path, fnv);
 745                 if (shmfd == NULL) {
 746                         /* Object does not yet exist, create it if requested. */
 747                         if (uap->flags & O_CREAT) {
 748 #ifdef MAC
 749                                 error = mac_posixshm_check_create(td->td_ucred,
 750                                     path);
 751                                 if (error == 0) {
 752 #endif
 753                                         shmfd = shm_alloc(td->td_ucred, cmode);
 754                                         shm_insert(path, fnv, shmfd);
 755 #ifdef MAC
 756                                 }
 757 #endif
 758                         } else {
 759                                 free(path, M_SHMFD);
 760                                 error = ENOENT;
 761                         }
 762                 } else {
 763                         /*
 764                          * Object already exists, obtain a new
 765                          * reference if requested and permitted.
 766                          */
 767                         free(path, M_SHMFD);
 768                         if ((uap->flags & (O_CREAT | O_EXCL)) ==
 769                             (O_CREAT | O_EXCL))
 770                                 error = EEXIST;
 771                         else {
 772 #ifdef MAC
 773                                 error = mac_posixshm_check_open(td->td_ucred,
 774                                     shmfd, FFLAGS(uap->flags & O_ACCMODE));
 775                                 if (error == 0)
 776 #endif
 777                                 error = shm_access(shmfd, td->td_ucred,
 778                                     FFLAGS(uap->flags & O_ACCMODE));
 779                         }
 780
 781                         /*
 782                          * Truncate the file back to zero length if
 783                          * O_TRUNC was specified and the object was
 784                          * opened with read/write.
 785                          */
 786                         if (error == 0 &&
 787                             (uap->flags & (O_ACCMODE | O_TRUNC)) ==
 788                             (O_RDWR | O_TRUNC)) {
 789 #ifdef MAC
 790                                 error = mac_posixshm_check_truncate(
 791                                         td->td_ucred, fp->f_cred, shmfd);
 792                                 if (error == 0)
 793 #endif
 794                                         shm_dotruncate(shmfd, 0);
 795                         }
 796                         if (error == 0)
 797                                 shm_hold(shmfd);
 798                 }
 799                 sx_xunlock(&shm_dict_lock);
 800
 801                 if (error) {
 802                         fdclose(fdp, fp, fd, td);
 803                         fdrop(fp, td);
 804                         return (error);
 805                 }
 806         }
 807
 808         finit(fp, FFLAGS(uap->flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops);
 809
 810         td->td_retval[0] = fd;
 811         fdrop(fp, td);
 812
 813         return (0);
 814 }
 815
 816 int
 817 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap)
 818 {
 819         char *path;
 820         Fnv32_t fnv;
 821         int error;
 822
 823         path = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
 824         error = copyinstr(uap->path, path, MAXPATHLEN, NULL);
 825         if (error) {
 826                 free(path, M_TEMP);
 827                 return (error);
 828         }
 829 #ifdef KTRACE
 830         if (KTRPOINT(curthread, KTR_NAMEI))
 831                 ktrnamei(path);
 832 #endif
 833         fnv = fnv_32_str(path, FNV1_32_INIT);
 834         sx_xlock(&shm_dict_lock);
 835         error = shm_remove(path, fnv, td->td_ucred);
 836         sx_xunlock(&shm_dict_lock);
 837         free(path, M_TEMP);
 838
 839         return (error);
 840 }
 841
 842 /*
 843  * mmap() helper to validate mmap() requests against shm object state
 844  * and give mmap() the vm_object to use for the mapping.
 845  */
 846 int
 847 shm_mmap(struct shmfd *shmfd, vm_size_t objsize, vm_ooffset_t foff,
 848     vm_object_t *obj)
 849 {
 850
 851         /*
 852          * XXXRW: This validation is probably insufficient, and subject to
 853          * sign errors.  It should be fixed.
 854          */
 855         if (foff >= shmfd->shm_size ||
 856             foff + objsize > round_page(shmfd->shm_size))
 857                 return (EINVAL);
 858
 859         mtx_lock(&shm_timestamp_lock);
 860         vfs_timestamp(&shmfd->shm_atime);
 861         mtx_unlock(&shm_timestamp_lock);
 862         vm_object_reference(shmfd->shm_object);
 863         *obj = shmfd->shm_object;
 864         return (0);
 865 }
 866
 867 static int
 868 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
 869     struct thread *td)
 870 {
 871         struct shmfd *shmfd;
 872         int error;
 873
 874         error = 0;
 875         shmfd = fp->f_data;
 876         mtx_lock(&shm_timestamp_lock);
 877         /*
 878          * SUSv4 says that x bits of permission need not be affected.
 879          * Be consistent with our shm_open there.
 880          */
 881 #ifdef MAC
 882         error = mac_posixshm_check_setmode(active_cred, shmfd, mode);
 883         if (error != 0)
 884                 goto out;
 885 #endif
 886         error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid,
 887             shmfd->shm_gid, VADMIN, active_cred, NULL);
 888         if (error != 0)
 889                 goto out;
 890         shmfd->shm_mode = mode & ACCESSPERMS;
 891 out:
 892         mtx_unlock(&shm_timestamp_lock);
 893         return (error);
 894 }
 895
 896 static int
 897 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
 898     struct thread *td)
 899 {
 900         struct shmfd *shmfd;
 901         int error;
 902
 903         error = 0;
 904         shmfd = fp->f_data;
 905         mtx_lock(&shm_timestamp_lock);
 906 #ifdef MAC
 907         error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid);
 908         if (error != 0)
 909                 goto out;
 910 #endif
 911         if (uid == (uid_t)-1)
 912                 uid = shmfd->shm_uid;
 913         if (gid == (gid_t)-1)
 914                  gid = shmfd->shm_gid;
 915         if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) ||
 916             (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) &&
 917             (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN, 0)))
 918                 goto out;
 919         shmfd->shm_uid = uid;
 920         shmfd->shm_gid = gid;
 921 out:
 922         mtx_unlock(&shm_timestamp_lock);
 923         return (error);
 924 }
 925
 926 /*
 927  * Helper routines to allow the backing object of a shared memory file
 928  * descriptor to be mapped in the kernel.
 929  */
 930 int
 931 shm_map(struct file *fp, size_t size, off_t offset, void **memp)
 932 {
 933         struct shmfd *shmfd;
 934         vm_offset_t kva, ofs;
 935         vm_object_t obj;
 936         int rv;
 937
 938         if (fp->f_type != DTYPE_SHM)
 939                 return (EINVAL);
 940         shmfd = fp->f_data;
 941         obj = shmfd->shm_object;
 942         VM_OBJECT_WLOCK(obj);
 943         /*
 944          * XXXRW: This validation is probably insufficient, and subject to
 945          * sign errors.  It should be fixed.
 946          */
 947         if (offset >= shmfd->shm_size ||
 948             offset + size > round_page(shmfd->shm_size)) {
 949                 VM_OBJECT_WUNLOCK(obj);
 950                 return (EINVAL);
 951         }
 952
 953         shmfd->shm_kmappings++;
 954         vm_object_reference_locked(obj);
 955         VM_OBJECT_WUNLOCK(obj);
 956
 957         /* Map the object into the kernel_map and wire it. */
 958         kva = vm_map_min(kernel_map);
 959         ofs = offset & PAGE_MASK;
 960         offset = trunc_page(offset);
 961         size = round_page(size + ofs);
 962         rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0,
 963             VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE,
 964             VM_PROT_READ | VM_PROT_WRITE, 0);
 965         if (rv == KERN_SUCCESS) {
 966                 rv = vm_map_wire(kernel_map, kva, kva + size,
 967                     VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
 968                 if (rv == KERN_SUCCESS) {
 969                         *memp = (void *)(kva + ofs);
 970                         return (0);
 971                 }
 972                 vm_map_remove(kernel_map, kva, kva + size);
 973         } else
 974                 vm_object_deallocate(obj);
 975
 976         /* On failure, drop our mapping reference. */
 977         VM_OBJECT_WLOCK(obj);
 978         shmfd->shm_kmappings--;
 979         VM_OBJECT_WUNLOCK(obj);
 980
 981         return (vm_mmap_to_errno(rv));
 982 }
 983
 984 /*
 985  * We require the caller to unmap the entire entry.  This allows us to
 986  * safely decrement shm_kmappings when a mapping is removed.
 987  */
 988 int
 989 shm_unmap(struct file *fp, void *mem, size_t size)
 990 {
 991         struct shmfd *shmfd;
 992         vm_map_entry_t entry;
 993         vm_offset_t kva, ofs;
 994         vm_object_t obj;
 995         vm_pindex_t pindex;
 996         vm_prot_t prot;
 997         boolean_t wired;
 998         vm_map_t map;
 999         int rv;
1000
1001         if (fp->f_type != DTYPE_SHM)
1002                 return (EINVAL);
1003         shmfd = fp->f_data;
1004         kva = (vm_offset_t)mem;
1005         ofs = kva & PAGE_MASK;
1006         kva = trunc_page(kva);
1007         size = round_page(size + ofs);
1008         map = kernel_map;
1009         rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry,
1010             &obj, &pindex, &prot, &wired);
1011         if (rv != KERN_SUCCESS)
1012                 return (EINVAL);
1013         if (entry->start != kva || entry->end != kva + size) {
1014                 vm_map_lookup_done(map, entry);
1015                 return (EINVAL);
1016         }
1017         vm_map_lookup_done(map, entry);
1018         if (obj != shmfd->shm_object)
1019                 return (EINVAL);
1020         vm_map_remove(map, kva, kva + size);
1021         VM_OBJECT_WLOCK(obj);
1022         KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped"));
1023         shmfd->shm_kmappings--;
1024         VM_OBJECT_WUNLOCK(obj);
1025         return (0);
1026 }
1027
1028 static int
1029 shm_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1030 {
1031         struct shmfd *shmfd;
1032
1033         kif->kf_type = KF_TYPE_SHM;
1034         shmfd = fp->f_data;
1035
1036         mtx_lock(&shm_timestamp_lock);
1037         kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode;    /* XXX */
1038         mtx_unlock(&shm_timestamp_lock);
1039         kif->kf_un.kf_file.kf_file_size = shmfd->shm_size;
1040         if (shmfd->shm_path != NULL) {
1041                 sx_slock(&shm_dict_lock);
1042                 if (shmfd->shm_path != NULL)
1043                         strlcpy(kif->kf_path, shmfd->shm_path,
1044                             sizeof(kif->kf_path));
1045                 sx_sunlock(&shm_dict_lock);
1046         }
1047         return (0);
1048 }