sys/kern/uipc_shm.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 2006, 2011, 2016-2017 Robert N. M. Watson
   5  * Copyright 2020 The FreeBSD Foundation
   6  * All rights reserved.
   7  *
   8  * Portions of this software were developed by BAE Systems, the University of
   9  * Cambridge Computer Laboratory, and Memorial University under DARPA/AFRL
  10  * contract FA8650-15-C-7558 ("CADETS"), as part of the DARPA Transparent
  11  * Computing (TC) research program.
  12  *
  13  * Portions of this software were developed by Konstantin Belousov
  14  * under sponsorship from the FreeBSD Foundation.
  15  *
  16  * Redistribution and use in source and binary forms, with or without
  17  * modification, are permitted provided that the following conditions
  18  * are met:
  19  * 1. Redistributions of source code must retain the above copyright
  20  *    notice, this list of conditions and the following disclaimer.
  21  * 2. Redistributions in binary form must reproduce the above copyright
  22  *    notice, this list of conditions and the following disclaimer in the
  23  *    documentation and/or other materials provided with the distribution.
  24  *
  25  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  26  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  27  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  28  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  29  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  30  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  31  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  32  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  33  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  34  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  35  * SUCH DAMAGE.
  36  */
  37
  38 /*
  39  * Support for shared swap-backed anonymous memory objects via
  40  * shm_open(2), shm_rename(2), and shm_unlink(2).
  41  * While most of the implementation is here, vm_mmap.c contains
  42  * mapping logic changes.
  43  *
  44  * posixshmcontrol(1) allows users to inspect the state of the memory
  45  * objects.  Per-uid swap resource limit controls total amount of
  46  * memory that user can consume for anonymous objects, including
  47  * shared.
  48  */
  49
  50 #include <sys/cdefs.h>
  51 __FBSDID("$FreeBSD$");
  52
  53 #include "opt_capsicum.h"
  54 #include "opt_ktrace.h"
  55
  56 #include <sys/param.h>
  57 #include <sys/capsicum.h>
  58 #include <sys/conf.h>
  59 #include <sys/fcntl.h>
  60 #include <sys/file.h>
  61 #include <sys/filedesc.h>
  62 #include <sys/filio.h>
  63 #include <sys/fnv_hash.h>
  64 #include <sys/kernel.h>
  65 #include <sys/limits.h>
  66 #include <sys/uio.h>
  67 #include <sys/signal.h>
  68 #include <sys/jail.h>
  69 #include <sys/ktrace.h>
  70 #include <sys/lock.h>
  71 #include <sys/malloc.h>
  72 #include <sys/mman.h>
  73 #include <sys/mutex.h>
  74 #include <sys/priv.h>
  75 #include <sys/proc.h>
  76 #include <sys/refcount.h>
  77 #include <sys/resourcevar.h>
  78 #include <sys/rwlock.h>
  79 #include <sys/sbuf.h>
  80 #include <sys/stat.h>
  81 #include <sys/syscallsubr.h>
  82 #include <sys/sysctl.h>
  83 #include <sys/sysproto.h>
  84 #include <sys/systm.h>
  85 #include <sys/sx.h>
  86 #include <sys/time.h>
  87 #include <sys/vmmeter.h>
  88 #include <sys/vnode.h>
  89 #include <sys/unistd.h>
  90 #include <sys/user.h>
  91
  92 #include <security/audit/audit.h>
  93 #include <security/mac/mac_framework.h>
  94
  95 #include <vm/vm.h>
  96 #include <vm/vm_param.h>
  97 #include <vm/pmap.h>
  98 #include <vm/vm_extern.h>
  99 #include <vm/vm_map.h>
 100 #include <vm/vm_kern.h>
 101 #include <vm/vm_object.h>
 102 #include <vm/vm_page.h>
 103 #include <vm/vm_pageout.h>
 104 #include <vm/vm_pager.h>
 105 #include <vm/swap_pager.h>
 106
 107 struct shm_mapping {
 108         char            *sm_path;
 109         Fnv32_t         sm_fnv;
 110         struct shmfd    *sm_shmfd;
 111         LIST_ENTRY(shm_mapping) sm_link;
 112 };
 113
 114 static MALLOC_DEFINE(M_SHMFD, "shmfd", "shared memory file descriptor");
 115 static LIST_HEAD(, shm_mapping) *shm_dictionary;
 116 static struct sx shm_dict_lock;
 117 static struct mtx shm_timestamp_lock;
 118 static u_long shm_hash;
 119 static struct unrhdr64 shm_ino_unr;
 120 static dev_t shm_dev_ino;
 121
 122 #define SHM_HASH(fnv)   (&shm_dictionary[(fnv) & shm_hash])
 123
 124 static void     shm_init(void *arg);
 125 static void     shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd);
 126 static struct shmfd *shm_lookup(char *path, Fnv32_t fnv);
 127 static int      shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred);
 128 static void     shm_doremove(struct shm_mapping *map);
 129 static int      shm_dotruncate_cookie(struct shmfd *shmfd, off_t length,
 130     void *rl_cookie);
 131 static int      shm_dotruncate_locked(struct shmfd *shmfd, off_t length,
 132     void *rl_cookie);
 133 static int      shm_copyin_path(struct thread *td, const char *userpath_in,
 134     char **path_out);
 135
 136 static fo_rdwr_t        shm_read;
 137 static fo_rdwr_t        shm_write;
 138 static fo_truncate_t    shm_truncate;
 139 static fo_ioctl_t       shm_ioctl;
 140 static fo_stat_t        shm_stat;
 141 static fo_close_t       shm_close;
 142 static fo_chmod_t       shm_chmod;
 143 static fo_chown_t       shm_chown;
 144 static fo_seek_t        shm_seek;
 145 static fo_fill_kinfo_t  shm_fill_kinfo;
 146 static fo_mmap_t        shm_mmap;
 147 static fo_get_seals_t   shm_get_seals;
 148 static fo_add_seals_t   shm_add_seals;
 149 static fo_fallocate_t   shm_fallocate;
 150
 151 /* File descriptor operations. */
 152 struct fileops shm_ops = {
 153         .fo_read = shm_read,
 154         .fo_write = shm_write,
 155         .fo_truncate = shm_truncate,
 156         .fo_ioctl = shm_ioctl,
 157         .fo_poll = invfo_poll,
 158         .fo_kqfilter = invfo_kqfilter,
 159         .fo_stat = shm_stat,
 160         .fo_close = shm_close,
 161         .fo_chmod = shm_chmod,
 162         .fo_chown = shm_chown,
 163         .fo_sendfile = vn_sendfile,
 164         .fo_seek = shm_seek,
 165         .fo_fill_kinfo = shm_fill_kinfo,
 166         .fo_mmap = shm_mmap,
 167         .fo_get_seals = shm_get_seals,
 168         .fo_add_seals = shm_add_seals,
 169         .fo_fallocate = shm_fallocate,
 170         .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE,
 171 };
 172
 173 FEATURE(posix_shm, "POSIX shared memory");
 174
 175 static SYSCTL_NODE(_vm, OID_AUTO, largepages, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
 176     "");
 177
 178 static int largepage_reclaim_tries = 1;
 179 SYSCTL_INT(_vm_largepages, OID_AUTO, reclaim_tries,
 180     CTLFLAG_RWTUN, &largepage_reclaim_tries, 0,
 181     "Number of contig reclaims before giving up for default alloc policy");
 182
 183 static int
 184 uiomove_object_page(vm_object_t obj, size_t len, struct uio *uio)
 185 {
 186         vm_page_t m;
 187         vm_pindex_t idx;
 188         size_t tlen;
 189         int error, offset, rv;
 190
 191         idx = OFF_TO_IDX(uio->uio_offset);
 192         offset = uio->uio_offset & PAGE_MASK;
 193         tlen = MIN(PAGE_SIZE - offset, len);
 194
 195         rv = vm_page_grab_valid_unlocked(&m, obj, idx,
 196             VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY | VM_ALLOC_NOCREAT);
 197         if (rv == VM_PAGER_OK)
 198                 goto found;
 199
 200         /*
 201          * Read I/O without either a corresponding resident page or swap
 202          * page: use zero_region.  This is intended to avoid instantiating
 203          * pages on read from a sparse region.
 204          */
 205         VM_OBJECT_WLOCK(obj);
 206         m = vm_page_lookup(obj, idx);
 207         if (uio->uio_rw == UIO_READ && m == NULL &&
 208             !vm_pager_has_page(obj, idx, NULL, NULL)) {
 209                 VM_OBJECT_WUNLOCK(obj);
 210                 return (uiomove(__DECONST(void *, zero_region), tlen, uio));
 211         }
 212
 213         /*
 214          * Although the tmpfs vnode lock is held here, it is
 215          * nonetheless safe to sleep waiting for a free page.  The
 216          * pageout daemon does not need to acquire the tmpfs vnode
 217          * lock to page out tobj's pages because tobj is a OBJT_SWAP
 218          * type object.
 219          */
 220         rv = vm_page_grab_valid(&m, obj, idx,
 221             VM_ALLOC_NORMAL | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY);
 222         if (rv != VM_PAGER_OK) {
 223                 VM_OBJECT_WUNLOCK(obj);
 224                 if (bootverbose) {
 225                         printf("uiomove_object: vm_obj %p idx %jd "
 226                             "pager error %d\n", obj, idx, rv);
 227                 }
 228                 return (rv == VM_PAGER_AGAIN ? ENOSPC : EIO);
 229         }
 230         VM_OBJECT_WUNLOCK(obj);
 231
 232 found:
 233         error = uiomove_fromphys(&m, offset, tlen, uio);
 234         if (uio->uio_rw == UIO_WRITE && error == 0)
 235                 vm_page_set_dirty(m);
 236         vm_page_activate(m);
 237         vm_page_sunbusy(m);
 238
 239         return (error);
 240 }
 241
 242 int
 243 uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio)
 244 {
 245         ssize_t resid;
 246         size_t len;
 247         int error;
 248
 249         error = 0;
 250         while ((resid = uio->uio_resid) > 0) {
 251                 if (obj_size <= uio->uio_offset)
 252                         break;
 253                 len = MIN(obj_size - uio->uio_offset, resid);
 254                 if (len == 0)
 255                         break;
 256                 error = uiomove_object_page(obj, len, uio);
 257                 if (error != 0 || resid == uio->uio_resid)
 258                         break;
 259         }
 260         return (error);
 261 }
 262
 263 static u_long count_largepages[MAXPAGESIZES];
 264
 265 static int
 266 shm_largepage_phys_populate(vm_object_t object, vm_pindex_t pidx,
 267     int fault_type, vm_prot_t max_prot, vm_pindex_t *first, vm_pindex_t *last)
 268 {
 269         vm_page_t m __diagused;
 270         int psind;
 271
 272         psind = object->un_pager.phys.data_val;
 273         if (psind == 0 || pidx >= object->size)
 274                 return (VM_PAGER_FAIL);
 275         *first = rounddown2(pidx, pagesizes[psind] / PAGE_SIZE);
 276
 277         /*
 278          * We only busy the first page in the superpage run.  It is
 279          * useless to busy whole run since we only remove full
 280          * superpage, and it takes too long to busy e.g. 512 * 512 ==
 281          * 262144 pages constituing 1G amd64 superage.
 282          */
 283         m = vm_page_grab(object, *first, VM_ALLOC_NORMAL | VM_ALLOC_NOCREAT);
 284         MPASS(m != NULL);
 285
 286         *last = *first + atop(pagesizes[psind]) - 1;
 287         return (VM_PAGER_OK);
 288 }
 289
 290 static boolean_t
 291 shm_largepage_phys_haspage(vm_object_t object, vm_pindex_t pindex,
 292     int *before, int *after)
 293 {
 294         int psind;
 295
 296         psind = object->un_pager.phys.data_val;
 297         if (psind == 0 || pindex >= object->size)
 298                 return (FALSE);
 299         if (before != NULL) {
 300                 *before = pindex - rounddown2(pindex, pagesizes[psind] /
 301                     PAGE_SIZE);
 302         }
 303         if (after != NULL) {
 304                 *after = roundup2(pindex, pagesizes[psind] / PAGE_SIZE) -
 305                     pindex;
 306         }
 307         return (TRUE);
 308 }
 309
 310 static void
 311 shm_largepage_phys_ctor(vm_object_t object, vm_prot_t prot,
 312     vm_ooffset_t foff, struct ucred *cred)
 313 {
 314 }
 315
 316 static void
 317 shm_largepage_phys_dtor(vm_object_t object)
 318 {
 319         int psind;
 320
 321         psind = object->un_pager.phys.data_val;
 322         if (psind != 0) {
 323                 atomic_subtract_long(&count_largepages[psind],
 324                     object->size / (pagesizes[psind] / PAGE_SIZE));
 325                 vm_wire_sub(object->size);
 326         } else {
 327                 KASSERT(object->size == 0,
 328                     ("largepage phys obj %p not initialized bit size %#jx > 0",
 329                     object, (uintmax_t)object->size));
 330         }
 331 }
 332
 333 static const struct phys_pager_ops shm_largepage_phys_ops = {
 334         .phys_pg_populate =     shm_largepage_phys_populate,
 335         .phys_pg_haspage =      shm_largepage_phys_haspage,
 336         .phys_pg_ctor =         shm_largepage_phys_ctor,
 337         .phys_pg_dtor =         shm_largepage_phys_dtor,
 338 };
 339
 340 bool
 341 shm_largepage(struct shmfd *shmfd)
 342 {
 343         return (shmfd->shm_object->type == OBJT_PHYS);
 344 }
 345
 346 static void
 347 shm_pager_freespace(vm_object_t obj, vm_pindex_t start, vm_size_t size)
 348 {
 349         struct shmfd *shm;
 350         vm_size_t c;
 351
 352         swap_pager_freespace(obj, start, size, &c);
 353         if (c == 0)
 354                 return;
 355
 356         shm = obj->un_pager.swp.swp_priv;
 357         if (shm == NULL)
 358                 return;
 359         KASSERT(shm->shm_pages >= c,
 360             ("shm %p pages %jd free %jd", shm,
 361             (uintmax_t)shm->shm_pages, (uintmax_t)c));
 362         shm->shm_pages -= c;
 363 }
 364
 365 static void
 366 shm_page_inserted(vm_object_t obj, vm_page_t m)
 367 {
 368         struct shmfd *shm;
 369
 370         shm = obj->un_pager.swp.swp_priv;
 371         if (shm == NULL)
 372                 return;
 373         if (!vm_pager_has_page(obj, m->pindex, NULL, NULL))
 374                 shm->shm_pages += 1;
 375 }
 376
 377 static void
 378 shm_page_removed(vm_object_t obj, vm_page_t m)
 379 {
 380         struct shmfd *shm;
 381
 382         shm = obj->un_pager.swp.swp_priv;
 383         if (shm == NULL)
 384                 return;
 385         if (!vm_pager_has_page(obj, m->pindex, NULL, NULL)) {
 386                 KASSERT(shm->shm_pages >= 1,
 387                     ("shm %p pages %jd free 1", shm,
 388                     (uintmax_t)shm->shm_pages));
 389                 shm->shm_pages -= 1;
 390         }
 391 }
 392
 393 static struct pagerops shm_swap_pager_ops = {
 394         .pgo_kvme_type = KVME_TYPE_SWAP,
 395         .pgo_freespace = shm_pager_freespace,
 396         .pgo_page_inserted = shm_page_inserted,
 397         .pgo_page_removed = shm_page_removed,
 398 };
 399 static int shmfd_pager_type = -1;
 400
 401 static int
 402 shm_seek(struct file *fp, off_t offset, int whence, struct thread *td)
 403 {
 404         struct shmfd *shmfd;
 405         off_t foffset;
 406         int error;
 407
 408         shmfd = fp->f_data;
 409         foffset = foffset_lock(fp, 0);
 410         error = 0;
 411         switch (whence) {
 412         case L_INCR:
 413                 if (foffset < 0 ||
 414                     (offset > 0 && foffset > OFF_MAX - offset)) {
 415                         error = EOVERFLOW;
 416                         break;
 417                 }
 418                 offset += foffset;
 419                 break;
 420         case L_XTND:
 421                 if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) {
 422                         error = EOVERFLOW;
 423                         break;
 424                 }
 425                 offset += shmfd->shm_size;
 426                 break;
 427         case L_SET:
 428                 break;
 429         default:
 430                 error = EINVAL;
 431         }
 432         if (error == 0) {
 433                 if (offset < 0 || offset > shmfd->shm_size)
 434                         error = EINVAL;
 435                 else
 436                         td->td_uretoff.tdu_off = offset;
 437         }
 438         foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
 439         return (error);
 440 }
 441
 442 static int
 443 shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
 444     int flags, struct thread *td)
 445 {
 446         struct shmfd *shmfd;
 447         void *rl_cookie;
 448         int error;
 449
 450         shmfd = fp->f_data;
 451 #ifdef MAC
 452         error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd);
 453         if (error)
 454                 return (error);
 455 #endif
 456         foffset_lock_uio(fp, uio, flags);
 457         rl_cookie = rangelock_rlock(&shmfd->shm_rl, uio->uio_offset,
 458             uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
 459         error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
 460         rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
 461         foffset_unlock_uio(fp, uio, flags);
 462         return (error);
 463 }
 464
 465 static int
 466 shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
 467     int flags, struct thread *td)
 468 {
 469         struct shmfd *shmfd;
 470         void *rl_cookie;
 471         int error;
 472         off_t size;
 473
 474         shmfd = fp->f_data;
 475 #ifdef MAC
 476         error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd);
 477         if (error)
 478                 return (error);
 479 #endif
 480         if (shm_largepage(shmfd) && shmfd->shm_lp_psind == 0)
 481                 return (EINVAL);
 482         foffset_lock_uio(fp, uio, flags);
 483         if (uio->uio_resid > OFF_MAX - uio->uio_offset) {
 484                 /*
 485                  * Overflow is only an error if we're supposed to expand on
 486                  * write.  Otherwise, we'll just truncate the write to the
 487                  * size of the file, which can only grow up to OFF_MAX.
 488                  */
 489                 if ((shmfd->shm_flags & SHM_GROW_ON_WRITE) != 0) {
 490                         foffset_unlock_uio(fp, uio, flags);
 491                         return (EFBIG);
 492                 }
 493
 494                 size = shmfd->shm_size;
 495         } else {
 496                 size = uio->uio_offset + uio->uio_resid;
 497         }
 498         if ((flags & FOF_OFFSET) == 0) {
 499                 rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
 500                     &shmfd->shm_mtx);
 501         } else {
 502                 rl_cookie = rangelock_wlock(&shmfd->shm_rl, uio->uio_offset,
 503                     size, &shmfd->shm_mtx);
 504         }
 505         if ((shmfd->shm_seals & F_SEAL_WRITE) != 0) {
 506                 error = EPERM;
 507         } else {
 508                 error = 0;
 509                 if ((shmfd->shm_flags & SHM_GROW_ON_WRITE) != 0 &&
 510                     size > shmfd->shm_size) {
 511                         error = shm_dotruncate_cookie(shmfd, size, rl_cookie);
 512                 }
 513                 if (error == 0)
 514                         error = uiomove_object(shmfd->shm_object,
 515                             shmfd->shm_size, uio);
 516         }
 517         rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
 518         foffset_unlock_uio(fp, uio, flags);
 519         return (error);
 520 }
 521
 522 static int
 523 shm_truncate(struct file *fp, off_t length, struct ucred *active_cred,
 524     struct thread *td)
 525 {
 526         struct shmfd *shmfd;
 527 #ifdef MAC
 528         int error;
 529 #endif
 530
 531         shmfd = fp->f_data;
 532 #ifdef MAC
 533         error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd);
 534         if (error)
 535                 return (error);
 536 #endif
 537         return (shm_dotruncate(shmfd, length));
 538 }
 539
 540 int
 541 shm_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
 542     struct thread *td)
 543 {
 544         struct shmfd *shmfd;
 545         struct shm_largepage_conf *conf;
 546         void *rl_cookie;
 547
 548         shmfd = fp->f_data;
 549         switch (com) {
 550         case FIONBIO:
 551         case FIOASYNC:
 552                 /*
 553                  * Allow fcntl(fd, F_SETFL, O_NONBLOCK) to work,
 554                  * just like it would on an unlinked regular file
 555                  */
 556                 return (0);
 557         case FIOSSHMLPGCNF:
 558                 if (!shm_largepage(shmfd))
 559                         return (ENOTTY);
 560                 conf = data;
 561                 if (shmfd->shm_lp_psind != 0 &&
 562                     conf->psind != shmfd->shm_lp_psind)
 563                         return (EINVAL);
 564                 if (conf->psind <= 0 || conf->psind >= MAXPAGESIZES ||
 565                     pagesizes[conf->psind] == 0)
 566                         return (EINVAL);
 567                 if (conf->alloc_policy != SHM_LARGEPAGE_ALLOC_DEFAULT &&
 568                     conf->alloc_policy != SHM_LARGEPAGE_ALLOC_NOWAIT &&
 569                     conf->alloc_policy != SHM_LARGEPAGE_ALLOC_HARD)
 570                         return (EINVAL);
 571
 572                 rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
 573                     &shmfd->shm_mtx);
 574                 shmfd->shm_lp_psind = conf->psind;
 575                 shmfd->shm_lp_alloc_policy = conf->alloc_policy;
 576                 shmfd->shm_object->un_pager.phys.data_val = conf->psind;
 577                 rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
 578                 return (0);
 579         case FIOGSHMLPGCNF:
 580                 if (!shm_largepage(shmfd))
 581                         return (ENOTTY);
 582                 conf = data;
 583                 rl_cookie = rangelock_rlock(&shmfd->shm_rl, 0, OFF_MAX,
 584                     &shmfd->shm_mtx);
 585                 conf->psind = shmfd->shm_lp_psind;
 586                 conf->alloc_policy = shmfd->shm_lp_alloc_policy;
 587                 rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
 588                 return (0);
 589         default:
 590                 return (ENOTTY);
 591         }
 592 }
 593
 594 static int
 595 shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred,
 596     struct thread *td)
 597 {
 598         struct shmfd *shmfd;
 599 #ifdef MAC
 600         int error;
 601 #endif
 602
 603         shmfd = fp->f_data;
 604
 605 #ifdef MAC
 606         error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd);
 607         if (error)
 608                 return (error);
 609 #endif
 610
 611         /*
 612          * Attempt to return sanish values for fstat() on a memory file
 613          * descriptor.
 614          */
 615         bzero(sb, sizeof(*sb));
 616         sb->st_blksize = PAGE_SIZE;
 617         sb->st_size = shmfd->shm_size;
 618         mtx_lock(&shm_timestamp_lock);
 619         sb->st_atim = shmfd->shm_atime;
 620         sb->st_ctim = shmfd->shm_ctime;
 621         sb->st_mtim = shmfd->shm_mtime;
 622         sb->st_birthtim = shmfd->shm_birthtime;
 623         sb->st_mode = S_IFREG | shmfd->shm_mode;                /* XXX */
 624         sb->st_uid = shmfd->shm_uid;
 625         sb->st_gid = shmfd->shm_gid;
 626         mtx_unlock(&shm_timestamp_lock);
 627         sb->st_dev = shm_dev_ino;
 628         sb->st_ino = shmfd->shm_ino;
 629         sb->st_nlink = shmfd->shm_object->ref_count;
 630         if (shm_largepage(shmfd)) {
 631                 sb->st_blocks = shmfd->shm_object->size /
 632                     (pagesizes[shmfd->shm_lp_psind] >> PAGE_SHIFT);
 633         } else {
 634                 sb->st_blocks = shmfd->shm_pages;
 635         }
 636
 637         return (0);
 638 }
 639
 640 static int
 641 shm_close(struct file *fp, struct thread *td)
 642 {
 643         struct shmfd *shmfd;
 644
 645         shmfd = fp->f_data;
 646         fp->f_data = NULL;
 647         shm_drop(shmfd);
 648
 649         return (0);
 650 }
 651
 652 static int
 653 shm_copyin_path(struct thread *td, const char *userpath_in, char **path_out) {
 654         int error;
 655         char *path;
 656         const char *pr_path;
 657         size_t pr_pathlen;
 658
 659         path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK);
 660         pr_path = td->td_ucred->cr_prison->pr_path;
 661
 662         /* Construct a full pathname for jailed callers. */
 663         pr_pathlen = strcmp(pr_path, "/") ==
 664             0 ? 0 : strlcpy(path, pr_path, MAXPATHLEN);
 665         error = copyinstr(userpath_in, path + pr_pathlen,
 666             MAXPATHLEN - pr_pathlen, NULL);
 667         if (error != 0)
 668                 goto out;
 669
 670 #ifdef KTRACE
 671         if (KTRPOINT(curthread, KTR_NAMEI))
 672                 ktrnamei(path);
 673 #endif
 674
 675         /* Require paths to start with a '/' character. */
 676         if (path[pr_pathlen] != '/') {
 677                 error = EINVAL;
 678                 goto out;
 679         }
 680
 681         *path_out = path;
 682
 683 out:
 684         if (error != 0)
 685                 free(path, M_SHMFD);
 686
 687         return (error);
 688 }
 689
 690 static int
 691 shm_dotruncate_locked(struct shmfd *shmfd, off_t length, void *rl_cookie)
 692 {
 693         vm_object_t object;
 694         vm_page_t m;
 695         vm_pindex_t idx, nobjsize;
 696         vm_ooffset_t delta;
 697         int base, rv;
 698
 699         KASSERT(length >= 0, ("shm_dotruncate: length < 0"));
 700         object = shmfd->shm_object;
 701         VM_OBJECT_ASSERT_WLOCKED(object);
 702         rangelock_cookie_assert(rl_cookie, RA_WLOCKED);
 703         if (length == shmfd->shm_size)
 704                 return (0);
 705         nobjsize = OFF_TO_IDX(length + PAGE_MASK);
 706
 707         /* Are we shrinking?  If so, trim the end. */
 708         if (length < shmfd->shm_size) {
 709                 if ((shmfd->shm_seals & F_SEAL_SHRINK) != 0)
 710                         return (EPERM);
 711
 712                 /*
 713                  * Disallow any requests to shrink the size if this
 714                  * object is mapped into the kernel.
 715                  */
 716                 if (shmfd->shm_kmappings > 0)
 717                         return (EBUSY);
 718
 719                 /*
 720                  * Zero the truncated part of the last page.
 721                  */
 722                 base = length & PAGE_MASK;
 723                 if (base != 0) {
 724                         idx = OFF_TO_IDX(length);
 725 retry:
 726                         m = vm_page_grab(object, idx, VM_ALLOC_NOCREAT);
 727                         if (m != NULL) {
 728                                 MPASS(vm_page_all_valid(m));
 729                         } else if (vm_pager_has_page(object, idx, NULL, NULL)) {
 730                                 m = vm_page_alloc(object, idx,
 731                                     VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL);
 732                                 if (m == NULL)
 733                                         goto retry;
 734                                 vm_object_pip_add(object, 1);
 735                                 VM_OBJECT_WUNLOCK(object);
 736                                 rv = vm_pager_get_pages(object, &m, 1, NULL,
 737                                     NULL);
 738                                 VM_OBJECT_WLOCK(object);
 739                                 vm_object_pip_wakeup(object);
 740                                 if (rv == VM_PAGER_OK) {
 741                                         /*
 742                                          * Since the page was not resident,
 743                                          * and therefore not recently
 744                                          * accessed, immediately enqueue it
 745                                          * for asynchronous laundering.  The
 746                                          * current operation is not regarded
 747                                          * as an access.
 748                                          */
 749                                         vm_page_launder(m);
 750                                 } else {
 751                                         vm_page_free(m);
 752                                         VM_OBJECT_WUNLOCK(object);
 753                                         return (EIO);
 754                                 }
 755                         }
 756                         if (m != NULL) {
 757                                 pmap_zero_page_area(m, base, PAGE_SIZE - base);
 758                                 KASSERT(vm_page_all_valid(m),
 759                                     ("shm_dotruncate: page %p is invalid", m));
 760                                 vm_page_set_dirty(m);
 761                                 vm_page_xunbusy(m);
 762                         }
 763                 }
 764                 delta = IDX_TO_OFF(object->size - nobjsize);
 765
 766                 if (nobjsize < object->size)
 767                         vm_object_page_remove(object, nobjsize, object->size,
 768                             0);
 769
 770                 /* Free the swap accounted for shm */
 771                 swap_release_by_cred(delta, object->cred);
 772                 object->charge -= delta;
 773         } else {
 774                 if ((shmfd->shm_seals & F_SEAL_GROW) != 0)
 775                         return (EPERM);
 776
 777                 /* Try to reserve additional swap space. */
 778                 delta = IDX_TO_OFF(nobjsize - object->size);
 779                 if (!swap_reserve_by_cred(delta, object->cred))
 780                         return (ENOMEM);
 781                 object->charge += delta;
 782         }
 783         shmfd->shm_size = length;
 784         mtx_lock(&shm_timestamp_lock);
 785         vfs_timestamp(&shmfd->shm_ctime);
 786         shmfd->shm_mtime = shmfd->shm_ctime;
 787         mtx_unlock(&shm_timestamp_lock);
 788         object->size = nobjsize;
 789         return (0);
 790 }
 791
 792 static int
 793 shm_dotruncate_largepage(struct shmfd *shmfd, off_t length, void *rl_cookie)
 794 {
 795         vm_object_t object;
 796         vm_page_t m;
 797         vm_pindex_t newobjsz;
 798         vm_pindex_t oldobjsz __unused;
 799         int aflags, error, i, psind, try;
 800
 801         KASSERT(length >= 0, ("shm_dotruncate: length < 0"));
 802         object = shmfd->shm_object;
 803         VM_OBJECT_ASSERT_WLOCKED(object);
 804         rangelock_cookie_assert(rl_cookie, RA_WLOCKED);
 805
 806         oldobjsz = object->size;
 807         newobjsz = OFF_TO_IDX(length);
 808         if (length == shmfd->shm_size)
 809                 return (0);
 810         psind = shmfd->shm_lp_psind;
 811         if (psind == 0 && length != 0)
 812                 return (EINVAL);
 813         if ((length & (pagesizes[psind] - 1)) != 0)
 814                 return (EINVAL);
 815
 816         if (length < shmfd->shm_size) {
 817                 if ((shmfd->shm_seals & F_SEAL_SHRINK) != 0)
 818                         return (EPERM);
 819                 if (shmfd->shm_kmappings > 0)
 820                         return (EBUSY);
 821                 return (ENOTSUP);       /* Pages are unmanaged. */
 822 #if 0
 823                 vm_object_page_remove(object, newobjsz, oldobjsz, 0);
 824                 object->size = newobjsz;
 825                 shmfd->shm_size = length;
 826                 return (0);
 827 #endif
 828         }
 829
 830         if ((shmfd->shm_seals & F_SEAL_GROW) != 0)
 831                 return (EPERM);
 832
 833         aflags = VM_ALLOC_NORMAL | VM_ALLOC_ZERO;
 834         if (shmfd->shm_lp_alloc_policy == SHM_LARGEPAGE_ALLOC_NOWAIT)
 835                 aflags |= VM_ALLOC_WAITFAIL;
 836         try = 0;
 837
 838         /*
 839          * Extend shmfd and object, keeping all already fully
 840          * allocated large pages intact even on error, because dropped
 841          * object lock might allowed mapping of them.
 842          */
 843         while (object->size < newobjsz) {
 844                 m = vm_page_alloc_contig(object, object->size, aflags,
 845                     pagesizes[psind] / PAGE_SIZE, 0, ~0,
 846                     pagesizes[psind], 0,
 847                     VM_MEMATTR_DEFAULT);
 848                 if (m == NULL) {
 849                         VM_OBJECT_WUNLOCK(object);
 850                         if (shmfd->shm_lp_alloc_policy ==
 851                             SHM_LARGEPAGE_ALLOC_NOWAIT ||
 852                             (shmfd->shm_lp_alloc_policy ==
 853                             SHM_LARGEPAGE_ALLOC_DEFAULT &&
 854                             try >= largepage_reclaim_tries)) {
 855                                 VM_OBJECT_WLOCK(object);
 856                                 return (ENOMEM);
 857                         }
 858                         error = vm_page_reclaim_contig(aflags,
 859                             pagesizes[psind] / PAGE_SIZE, 0, ~0,
 860                             pagesizes[psind], 0) ? 0 :
 861                             vm_wait_intr(object);
 862                         if (error != 0) {
 863                                 VM_OBJECT_WLOCK(object);
 864                                 return (error);
 865                         }
 866                         try++;
 867                         VM_OBJECT_WLOCK(object);
 868                         continue;
 869                 }
 870                 try = 0;
 871                 for (i = 0; i < pagesizes[psind] / PAGE_SIZE; i++) {
 872                         if ((m[i].flags & PG_ZERO) == 0)
 873                                 pmap_zero_page(&m[i]);
 874                         vm_page_valid(&m[i]);
 875                         vm_page_xunbusy(&m[i]);
 876                 }
 877                 object->size += OFF_TO_IDX(pagesizes[psind]);
 878                 shmfd->shm_size += pagesizes[psind];
 879                 atomic_add_long(&count_largepages[psind], 1);
 880                 vm_wire_add(atop(pagesizes[psind]));
 881         }
 882         return (0);
 883 }
 884
 885 static int
 886 shm_dotruncate_cookie(struct shmfd *shmfd, off_t length, void *rl_cookie)
 887 {
 888         int error;
 889
 890         VM_OBJECT_WLOCK(shmfd->shm_object);
 891         error = shm_largepage(shmfd) ? shm_dotruncate_largepage(shmfd,
 892             length, rl_cookie) : shm_dotruncate_locked(shmfd, length,
 893             rl_cookie);
 894         VM_OBJECT_WUNLOCK(shmfd->shm_object);
 895         return (error);
 896 }
 897
 898 int
 899 shm_dotruncate(struct shmfd *shmfd, off_t length)
 900 {
 901         void *rl_cookie;
 902         int error;
 903
 904         rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
 905             &shmfd->shm_mtx);
 906         error = shm_dotruncate_cookie(shmfd, length, rl_cookie);
 907         rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
 908         return (error);
 909 }
 910
 911 /*
 912  * shmfd object management including creation and reference counting
 913  * routines.
 914  */
 915 struct shmfd *
 916 shm_alloc(struct ucred *ucred, mode_t mode, bool largepage)
 917 {
 918         struct shmfd *shmfd;
 919         vm_object_t obj;
 920
 921         shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO);
 922         shmfd->shm_size = 0;
 923         shmfd->shm_uid = ucred->cr_uid;
 924         shmfd->shm_gid = ucred->cr_gid;
 925         shmfd->shm_mode = mode;
 926         if (largepage) {
 927                 shmfd->shm_object = phys_pager_allocate(NULL,
 928                     &shm_largepage_phys_ops, NULL, shmfd->shm_size,
 929                     VM_PROT_DEFAULT, 0, ucred);
 930                 shmfd->shm_lp_alloc_policy = SHM_LARGEPAGE_ALLOC_DEFAULT;
 931         } else {
 932                 obj = vm_pager_allocate(shmfd_pager_type, NULL,
 933                     shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred);
 934                 VM_OBJECT_WLOCK(obj);
 935                 obj->un_pager.swp.swp_priv = shmfd;
 936                 VM_OBJECT_WUNLOCK(obj);
 937                 shmfd->shm_object = obj;
 938         }
 939         KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate"));
 940         vfs_timestamp(&shmfd->shm_birthtime);
 941         shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime =
 942             shmfd->shm_birthtime;
 943         shmfd->shm_ino = alloc_unr64(&shm_ino_unr);
 944         refcount_init(&shmfd->shm_refs, 1);
 945         mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF);
 946         rangelock_init(&shmfd->shm_rl);
 947 #ifdef MAC
 948         mac_posixshm_init(shmfd);
 949         mac_posixshm_create(ucred, shmfd);
 950 #endif
 951
 952         return (shmfd);
 953 }
 954
 955 struct shmfd *
 956 shm_hold(struct shmfd *shmfd)
 957 {
 958
 959         refcount_acquire(&shmfd->shm_refs);
 960         return (shmfd);
 961 }
 962
 963 void
 964 shm_drop(struct shmfd *shmfd)
 965 {
 966         vm_object_t obj;
 967
 968         if (refcount_release(&shmfd->shm_refs)) {
 969 #ifdef MAC
 970                 mac_posixshm_destroy(shmfd);
 971 #endif
 972                 rangelock_destroy(&shmfd->shm_rl);
 973                 mtx_destroy(&shmfd->shm_mtx);
 974                 obj = shmfd->shm_object;
 975                 if (!shm_largepage(shmfd)) {
 976                         VM_OBJECT_WLOCK(obj);
 977                         obj->un_pager.swp.swp_priv = NULL;
 978                         VM_OBJECT_WUNLOCK(obj);
 979                 }
 980                 vm_object_deallocate(obj);
 981                 free(shmfd, M_SHMFD);
 982         }
 983 }
 984
 985 /*
 986  * Determine if the credentials have sufficient permissions for a
 987  * specified combination of FREAD and FWRITE.
 988  */
 989 int
 990 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags)
 991 {
 992         accmode_t accmode;
 993         int error;
 994
 995         accmode = 0;
 996         if (flags & FREAD)
 997                 accmode |= VREAD;
 998         if (flags & FWRITE)
 999                 accmode |= VWRITE;
1000         mtx_lock(&shm_timestamp_lock);
1001         error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
1002             accmode, ucred);
1003         mtx_unlock(&shm_timestamp_lock);
1004         return (error);
1005 }
1006
1007 static void
1008 shm_init(void *arg)
1009 {
1010         char name[32];
1011         int i;
1012
1013         mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF);
1014         sx_init(&shm_dict_lock, "shm dictionary");
1015         shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash);
1016         new_unrhdr64(&shm_ino_unr, 1);
1017         shm_dev_ino = devfs_alloc_cdp_inode();
1018         KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized"));
1019         shmfd_pager_type = vm_pager_alloc_dyn_type(&shm_swap_pager_ops,
1020             OBJT_SWAP);
1021         MPASS(shmfd_pager_type != -1);
1022
1023         for (i = 1; i < MAXPAGESIZES; i++) {
1024                 if (pagesizes[i] == 0)
1025                         break;
1026 #define M       (1024 * 1024)
1027 #define G       (1024 * M)
1028                 if (pagesizes[i] >= G)
1029                         snprintf(name, sizeof(name), "%luG", pagesizes[i] / G);
1030                 else if (pagesizes[i] >= M)
1031                         snprintf(name, sizeof(name), "%luM", pagesizes[i] / M);
1032                 else
1033                         snprintf(name, sizeof(name), "%lu", pagesizes[i]);
1034 #undef G
1035 #undef M
1036                 SYSCTL_ADD_ULONG(NULL, SYSCTL_STATIC_CHILDREN(_vm_largepages),
1037                     OID_AUTO, name, CTLFLAG_RD, &count_largepages[i],
1038                     "number of non-transient largepages allocated");
1039         }
1040 }
1041 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL);
1042
1043 /*
1044  * Remove all shared memory objects that belong to a prison.
1045  */
1046 void
1047 shm_remove_prison(struct prison *pr)
1048 {
1049         struct shm_mapping *shmm, *tshmm;
1050         u_long i;
1051
1052         sx_xlock(&shm_dict_lock);
1053         for (i = 0; i < shm_hash + 1; i++) {
1054                 LIST_FOREACH_SAFE(shmm, &shm_dictionary[i], sm_link, tshmm) {
1055                         if (shmm->sm_shmfd->shm_object->cred &&
1056                             shmm->sm_shmfd->shm_object->cred->cr_prison == pr)
1057                                 shm_doremove(shmm);
1058                 }
1059         }
1060         sx_xunlock(&shm_dict_lock);
1061 }
1062
1063 /*
1064  * Dictionary management.  We maintain an in-kernel dictionary to map
1065  * paths to shmfd objects.  We use the FNV hash on the path to store
1066  * the mappings in a hash table.
1067  */
1068 static struct shmfd *
1069 shm_lookup(char *path, Fnv32_t fnv)
1070 {
1071         struct shm_mapping *map;
1072
1073         LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
1074                 if (map->sm_fnv != fnv)
1075                         continue;
1076                 if (strcmp(map->sm_path, path) == 0)
1077                         return (map->sm_shmfd);
1078         }
1079
1080         return (NULL);
1081 }
1082
1083 static void
1084 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd)
1085 {
1086         struct shm_mapping *map;
1087
1088         map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK);
1089         map->sm_path = path;
1090         map->sm_fnv = fnv;
1091         map->sm_shmfd = shm_hold(shmfd);
1092         shmfd->shm_path = path;
1093         LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link);
1094 }
1095
1096 static int
1097 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred)
1098 {
1099         struct shm_mapping *map;
1100         int error;
1101
1102         LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
1103                 if (map->sm_fnv != fnv)
1104                         continue;
1105                 if (strcmp(map->sm_path, path) == 0) {
1106 #ifdef MAC
1107                         error = mac_posixshm_check_unlink(ucred, map->sm_shmfd);
1108                         if (error)
1109                                 return (error);
1110 #endif
1111                         error = shm_access(map->sm_shmfd, ucred,
1112                             FREAD | FWRITE);
1113                         if (error)
1114                                 return (error);
1115                         shm_doremove(map);
1116                         return (0);
1117                 }
1118         }
1119
1120         return (ENOENT);
1121 }
1122
1123 static void
1124 shm_doremove(struct shm_mapping *map)
1125 {
1126         map->sm_shmfd->shm_path = NULL;
1127         LIST_REMOVE(map, sm_link);
1128         shm_drop(map->sm_shmfd);
1129         free(map->sm_path, M_SHMFD);
1130         free(map, M_SHMFD);
1131 }
1132
1133 int
1134 kern_shm_open2(struct thread *td, const char *userpath, int flags, mode_t mode,
1135     int shmflags, struct filecaps *fcaps, const char *name __unused)
1136 {
1137         struct pwddesc *pdp;
1138         struct shmfd *shmfd;
1139         struct file *fp;
1140         char *path;
1141         void *rl_cookie;
1142         Fnv32_t fnv;
1143         mode_t cmode;
1144         int error, fd, initial_seals;
1145         bool largepage;
1146
1147         if ((shmflags & ~(SHM_ALLOW_SEALING | SHM_GROW_ON_WRITE |
1148             SHM_LARGEPAGE)) != 0)
1149                 return (EINVAL);
1150
1151         initial_seals = F_SEAL_SEAL;
1152         if ((shmflags & SHM_ALLOW_SEALING) != 0)
1153                 initial_seals &= ~F_SEAL_SEAL;
1154
1155 #ifdef CAPABILITY_MODE
1156         /*
1157          * shm_open(2) is only allowed for anonymous objects.
1158          */
1159         if (IN_CAPABILITY_MODE(td) && (userpath != SHM_ANON))
1160                 return (ECAPMODE);
1161 #endif
1162
1163         AUDIT_ARG_FFLAGS(flags);
1164         AUDIT_ARG_MODE(mode);
1165
1166         if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR)
1167                 return (EINVAL);
1168
1169         if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0)
1170                 return (EINVAL);
1171
1172         largepage = (shmflags & SHM_LARGEPAGE) != 0;
1173         if (largepage && !PMAP_HAS_LARGEPAGES)
1174                 return (ENOTTY);
1175
1176         /*
1177          * Currently only F_SEAL_SEAL may be set when creating or opening shmfd.
1178          * If the decision is made later to allow additional seals, care must be
1179          * taken below to ensure that the seals are properly set if the shmfd
1180          * already existed -- this currently assumes that only F_SEAL_SEAL can
1181          * be set and doesn't take further precautions to ensure the validity of
1182          * the seals being added with respect to current mappings.
1183          */
1184         if ((initial_seals & ~F_SEAL_SEAL) != 0)
1185                 return (EINVAL);
1186
1187         pdp = td->td_proc->p_pd;
1188         cmode = (mode & ~pdp->pd_cmask) & ACCESSPERMS;
1189
1190         /*
1191          * shm_open(2) created shm should always have O_CLOEXEC set, as mandated
1192          * by POSIX.  We allow it to be unset here so that an in-kernel
1193          * interface may be written as a thin layer around shm, optionally not
1194          * setting CLOEXEC.  For shm_open(2), O_CLOEXEC is set unconditionally
1195          * in sys_shm_open() to keep this implementation compliant.
1196          */
1197         error = falloc_caps(td, &fp, &fd, flags & O_CLOEXEC, fcaps);
1198         if (error)
1199                 return (error);
1200
1201         /* A SHM_ANON path pointer creates an anonymous object. */
1202         if (userpath == SHM_ANON) {
1203                 /* A read-only anonymous object is pointless. */
1204                 if ((flags & O_ACCMODE) == O_RDONLY) {
1205                         fdclose(td, fp, fd);
1206                         fdrop(fp, td);
1207                         return (EINVAL);
1208                 }
1209                 shmfd = shm_alloc(td->td_ucred, cmode, largepage);
1210                 shmfd->shm_seals = initial_seals;
1211                 shmfd->shm_flags = shmflags;
1212         } else {
1213                 error = shm_copyin_path(td, userpath, &path);
1214                 if (error != 0) {
1215                         fdclose(td, fp, fd);
1216                         fdrop(fp, td);
1217                         return (error);
1218                 }
1219
1220                 AUDIT_ARG_UPATH1_CANON(path);
1221                 fnv = fnv_32_str(path, FNV1_32_INIT);
1222                 sx_xlock(&shm_dict_lock);
1223                 shmfd = shm_lookup(path, fnv);
1224                 if (shmfd == NULL) {
1225                         /* Object does not yet exist, create it if requested. */
1226                         if (flags & O_CREAT) {
1227 #ifdef MAC
1228                                 error = mac_posixshm_check_create(td->td_ucred,
1229                                     path);
1230                                 if (error == 0) {
1231 #endif
1232                                         shmfd = shm_alloc(td->td_ucred, cmode,
1233                                             largepage);
1234                                         shmfd->shm_seals = initial_seals;
1235                                         shmfd->shm_flags = shmflags;
1236                                         shm_insert(path, fnv, shmfd);
1237 #ifdef MAC
1238                                 }
1239 #endif
1240                         } else {
1241                                 free(path, M_SHMFD);
1242                                 error = ENOENT;
1243                         }
1244                 } else {
1245                         rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
1246                             &shmfd->shm_mtx);
1247
1248                         /*
1249                          * kern_shm_open() likely shouldn't ever error out on
1250                          * trying to set a seal that already exists, unlike
1251                          * F_ADD_SEALS.  This would break terribly as
1252                          * shm_open(2) actually sets F_SEAL_SEAL to maintain
1253                          * historical behavior where the underlying file could
1254                          * not be sealed.
1255                          */
1256                         initial_seals &= ~shmfd->shm_seals;
1257
1258                         /*
1259                          * Object already exists, obtain a new
1260                          * reference if requested and permitted.
1261                          */
1262                         free(path, M_SHMFD);
1263
1264                         /*
1265                          * initial_seals can't set additional seals if we've
1266                          * already been set F_SEAL_SEAL.  If F_SEAL_SEAL is set,
1267                          * then we've already removed that one from
1268                          * initial_seals.  This is currently redundant as we
1269                          * only allow setting F_SEAL_SEAL at creation time, but
1270                          * it's cheap to check and decreases the effort required
1271                          * to allow additional seals.
1272                          */
1273                         if ((shmfd->shm_seals & F_SEAL_SEAL) != 0 &&
1274                             initial_seals != 0)
1275                                 error = EPERM;
1276                         else if ((flags & (O_CREAT | O_EXCL)) ==
1277                             (O_CREAT | O_EXCL))
1278                                 error = EEXIST;
1279                         else if (shmflags != 0 && shmflags != shmfd->shm_flags)
1280                                 error = EINVAL;
1281                         else {
1282 #ifdef MAC
1283                                 error = mac_posixshm_check_open(td->td_ucred,
1284                                     shmfd, FFLAGS(flags & O_ACCMODE));
1285                                 if (error == 0)
1286 #endif
1287                                 error = shm_access(shmfd, td->td_ucred,
1288                                     FFLAGS(flags & O_ACCMODE));
1289                         }
1290
1291                         /*
1292                          * Truncate the file back to zero length if
1293                          * O_TRUNC was specified and the object was
1294                          * opened with read/write.
1295                          */
1296                         if (error == 0 &&
1297                             (flags & (O_ACCMODE | O_TRUNC)) ==
1298                             (O_RDWR | O_TRUNC)) {
1299                                 VM_OBJECT_WLOCK(shmfd->shm_object);
1300 #ifdef MAC
1301                                 error = mac_posixshm_check_truncate(
1302                                         td->td_ucred, fp->f_cred, shmfd);
1303                                 if (error == 0)
1304 #endif
1305                                         error = shm_dotruncate_locked(shmfd, 0,
1306                                             rl_cookie);
1307                                 VM_OBJECT_WUNLOCK(shmfd->shm_object);
1308                         }
1309                         if (error == 0) {
1310                                 /*
1311                                  * Currently we only allow F_SEAL_SEAL to be
1312                                  * set initially.  As noted above, this would
1313                                  * need to be reworked should that change.
1314                                  */
1315                                 shmfd->shm_seals |= initial_seals;
1316                                 shm_hold(shmfd);
1317                         }
1318                         rangelock_unlock(&shmfd->shm_rl, rl_cookie,
1319                             &shmfd->shm_mtx);
1320                 }
1321                 sx_xunlock(&shm_dict_lock);
1322
1323                 if (error) {
1324                         fdclose(td, fp, fd);
1325                         fdrop(fp, td);
1326                         return (error);
1327                 }
1328         }
1329
1330         finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops);
1331
1332         td->td_retval[0] = fd;
1333         fdrop(fp, td);
1334
1335         return (0);
1336 }
1337
1338 /* System calls. */
1339 #ifdef COMPAT_FREEBSD12
1340 int
1341 freebsd12_shm_open(struct thread *td, struct freebsd12_shm_open_args *uap)
1342 {
1343
1344         return (kern_shm_open(td, uap->path, uap->flags | O_CLOEXEC,
1345             uap->mode, NULL));
1346 }
1347 #endif
1348
1349 int
1350 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap)
1351 {
1352         char *path;
1353         Fnv32_t fnv;
1354         int error;
1355
1356         error = shm_copyin_path(td, uap->path, &path);
1357         if (error != 0)
1358                 return (error);
1359
1360         AUDIT_ARG_UPATH1_CANON(path);
1361         fnv = fnv_32_str(path, FNV1_32_INIT);
1362         sx_xlock(&shm_dict_lock);
1363         error = shm_remove(path, fnv, td->td_ucred);
1364         sx_xunlock(&shm_dict_lock);
1365         free(path, M_SHMFD);
1366
1367         return (error);
1368 }
1369
1370 int
1371 sys_shm_rename(struct thread *td, struct shm_rename_args *uap)
1372 {
1373         char *path_from = NULL, *path_to = NULL;
1374         Fnv32_t fnv_from, fnv_to;
1375         struct shmfd *fd_from;
1376         struct shmfd *fd_to;
1377         int error;
1378         int flags;
1379
1380         flags = uap->flags;
1381         AUDIT_ARG_FFLAGS(flags);
1382
1383         /*
1384          * Make sure the user passed only valid flags.
1385          * If you add a new flag, please add a new term here.
1386          */
1387         if ((flags & ~(
1388             SHM_RENAME_NOREPLACE |
1389             SHM_RENAME_EXCHANGE
1390             )) != 0) {
1391                 error = EINVAL;
1392                 goto out;
1393         }
1394
1395         /*
1396          * EXCHANGE and NOREPLACE don't quite make sense together. Let's
1397          * force the user to choose one or the other.
1398          */
1399         if ((flags & SHM_RENAME_NOREPLACE) != 0 &&
1400             (flags & SHM_RENAME_EXCHANGE) != 0) {
1401                 error = EINVAL;
1402                 goto out;
1403         }
1404
1405         /* Renaming to or from anonymous makes no sense */
1406         if (uap->path_from == SHM_ANON || uap->path_to == SHM_ANON) {
1407                 error = EINVAL;
1408                 goto out;
1409         }
1410
1411         error = shm_copyin_path(td, uap->path_from, &path_from);
1412         if (error != 0)
1413                 goto out;
1414
1415         error = shm_copyin_path(td, uap->path_to, &path_to);
1416         if (error != 0)
1417                 goto out;
1418
1419         AUDIT_ARG_UPATH1_CANON(path_from);
1420         AUDIT_ARG_UPATH2_CANON(path_to);
1421
1422         /* Rename with from/to equal is a no-op */
1423         if (strcmp(path_from, path_to) == 0)
1424                 goto out;
1425
1426         fnv_from = fnv_32_str(path_from, FNV1_32_INIT);
1427         fnv_to = fnv_32_str(path_to, FNV1_32_INIT);
1428
1429         sx_xlock(&shm_dict_lock);
1430
1431         fd_from = shm_lookup(path_from, fnv_from);
1432         if (fd_from == NULL) {
1433                 error = ENOENT;
1434                 goto out_locked;
1435         }
1436
1437         fd_to = shm_lookup(path_to, fnv_to);
1438         if ((flags & SHM_RENAME_NOREPLACE) != 0 && fd_to != NULL) {
1439                 error = EEXIST;
1440                 goto out_locked;
1441         }
1442
1443         /*
1444          * Unconditionally prevents shm_remove from invalidating the 'from'
1445          * shm's state.
1446          */
1447         shm_hold(fd_from);
1448         error = shm_remove(path_from, fnv_from, td->td_ucred);
1449
1450         /*
1451          * One of my assumptions failed if ENOENT (e.g. locking didn't
1452          * protect us)
1453          */
1454         KASSERT(error != ENOENT, ("Our shm disappeared during shm_rename: %s",
1455             path_from));
1456         if (error != 0) {
1457                 shm_drop(fd_from);
1458                 goto out_locked;
1459         }
1460
1461         /*
1462          * If we are exchanging, we need to ensure the shm_remove below
1463          * doesn't invalidate the dest shm's state.
1464          */
1465         if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL)
1466                 shm_hold(fd_to);
1467
1468         /*
1469          * NOTE: if path_to is not already in the hash, c'est la vie;
1470          * it simply means we have nothing already at path_to to unlink.
1471          * That is the ENOENT case.
1472          *
1473          * If we somehow don't have access to unlink this guy, but
1474          * did for the shm at path_from, then relink the shm to path_from
1475          * and abort with EACCES.
1476          *
1477          * All other errors: that is weird; let's relink and abort the
1478          * operation.
1479          */
1480         error = shm_remove(path_to, fnv_to, td->td_ucred);
1481         if (error != 0 && error != ENOENT) {
1482                 shm_insert(path_from, fnv_from, fd_from);
1483                 shm_drop(fd_from);
1484                 /* Don't free path_from now, since the hash references it */
1485                 path_from = NULL;
1486                 goto out_locked;
1487         }
1488
1489         error = 0;
1490
1491         shm_insert(path_to, fnv_to, fd_from);
1492
1493         /* Don't free path_to now, since the hash references it */
1494         path_to = NULL;
1495
1496         /* We kept a ref when we removed, and incremented again in insert */
1497         shm_drop(fd_from);
1498         KASSERT(fd_from->shm_refs > 0, ("Expected >0 refs; got: %d\n",
1499             fd_from->shm_refs));
1500
1501         if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) {
1502                 shm_insert(path_from, fnv_from, fd_to);
1503                 path_from = NULL;
1504                 shm_drop(fd_to);
1505                 KASSERT(fd_to->shm_refs > 0, ("Expected >0 refs; got: %d\n",
1506                     fd_to->shm_refs));
1507         }
1508
1509 out_locked:
1510         sx_xunlock(&shm_dict_lock);
1511
1512 out:
1513         free(path_from, M_SHMFD);
1514         free(path_to, M_SHMFD);
1515         return (error);
1516 }
1517
1518 static int
1519 shm_mmap_large(struct shmfd *shmfd, vm_map_t map, vm_offset_t *addr,
1520     vm_size_t size, vm_prot_t prot, vm_prot_t max_prot, int flags,
1521     vm_ooffset_t foff, struct thread *td)
1522 {
1523         struct vmspace *vms;
1524         vm_map_entry_t next_entry, prev_entry;
1525         vm_offset_t align, mask, maxaddr;
1526         int docow, error, rv, try;
1527         bool curmap;
1528
1529         if (shmfd->shm_lp_psind == 0)
1530                 return (EINVAL);
1531
1532         /* MAP_PRIVATE is disabled */
1533         if ((flags & ~(MAP_SHARED | MAP_FIXED | MAP_EXCL |
1534             MAP_NOCORE |
1535 #ifdef MAP_32BIT
1536             MAP_32BIT |
1537 #endif
1538             MAP_ALIGNMENT_MASK)) != 0)
1539                 return (EINVAL);
1540
1541         vms = td->td_proc->p_vmspace;
1542         curmap = map == &vms->vm_map;
1543         if (curmap) {
1544                 error = kern_mmap_racct_check(td, map, size);
1545                 if (error != 0)
1546                         return (error);
1547         }
1548
1549         docow = shmfd->shm_lp_psind << MAP_SPLIT_BOUNDARY_SHIFT;
1550         docow |= MAP_INHERIT_SHARE;
1551         if ((flags & MAP_NOCORE) != 0)
1552                 docow |= MAP_DISABLE_COREDUMP;
1553
1554         mask = pagesizes[shmfd->shm_lp_psind] - 1;
1555         if ((foff & mask) != 0)
1556                 return (EINVAL);
1557         maxaddr = vm_map_max(map);
1558 #ifdef MAP_32BIT
1559         if ((flags & MAP_32BIT) != 0 && maxaddr > MAP_32BIT_MAX_ADDR)
1560                 maxaddr = MAP_32BIT_MAX_ADDR;
1561 #endif
1562         if (size == 0 || (size & mask) != 0 ||
1563             (*addr != 0 && ((*addr & mask) != 0 ||
1564             *addr + size < *addr || *addr + size > maxaddr)))
1565                 return (EINVAL);
1566
1567         align = flags & MAP_ALIGNMENT_MASK;
1568         if (align == 0) {
1569                 align = pagesizes[shmfd->shm_lp_psind];
1570         } else if (align == MAP_ALIGNED_SUPER) {
1571                 if (shmfd->shm_lp_psind != 1)
1572                         return (EINVAL);
1573                 align = pagesizes[1];
1574         } else {
1575                 align >>= MAP_ALIGNMENT_SHIFT;
1576                 align = 1ULL << align;
1577                 /* Also handles overflow. */
1578                 if (align < pagesizes[shmfd->shm_lp_psind])
1579                         return (EINVAL);
1580         }
1581
1582         vm_map_lock(map);
1583         if ((flags & MAP_FIXED) == 0) {
1584                 try = 1;
1585                 if (curmap && (*addr == 0 ||
1586                     (*addr >= round_page((vm_offset_t)vms->vm_taddr) &&
1587                     *addr < round_page((vm_offset_t)vms->vm_daddr +
1588                     lim_max(td, RLIMIT_DATA))))) {
1589                         *addr = roundup2((vm_offset_t)vms->vm_daddr +
1590                             lim_max(td, RLIMIT_DATA),
1591                             pagesizes[shmfd->shm_lp_psind]);
1592                 }
1593 again:
1594                 rv = vm_map_find_aligned(map, addr, size, maxaddr, align);
1595                 if (rv != KERN_SUCCESS) {
1596                         if (try == 1) {
1597                                 try = 2;
1598                                 *addr = vm_map_min(map);
1599                                 if ((*addr & mask) != 0)
1600                                         *addr = (*addr + mask) & mask;
1601                                 goto again;
1602                         }
1603                         goto fail1;
1604                 }
1605         } else if ((flags & MAP_EXCL) == 0) {
1606                 rv = vm_map_delete(map, *addr, *addr + size);
1607                 if (rv != KERN_SUCCESS)
1608                         goto fail1;
1609         } else {
1610                 error = ENOSPC;
1611                 if (vm_map_lookup_entry(map, *addr, &prev_entry))
1612                         goto fail;
1613                 next_entry = vm_map_entry_succ(prev_entry);
1614                 if (next_entry->start < *addr + size)
1615                         goto fail;
1616         }
1617
1618         rv = vm_map_insert(map, shmfd->shm_object, foff, *addr, *addr + size,
1619             prot, max_prot, docow);
1620 fail1:
1621         error = vm_mmap_to_errno(rv);
1622 fail:
1623         vm_map_unlock(map);
1624         return (error);
1625 }
1626
1627 static int
1628 shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize,
1629     vm_prot_t prot, vm_prot_t cap_maxprot, int flags,
1630     vm_ooffset_t foff, struct thread *td)
1631 {
1632         struct shmfd *shmfd;
1633         vm_prot_t maxprot;
1634         int error;
1635         bool writecnt;
1636         void *rl_cookie;
1637
1638         shmfd = fp->f_data;
1639         maxprot = VM_PROT_NONE;
1640
1641         rl_cookie = rangelock_rlock(&shmfd->shm_rl, 0, objsize,
1642             &shmfd->shm_mtx);
1643         /* FREAD should always be set. */
1644         if ((fp->f_flag & FREAD) != 0)
1645                 maxprot |= VM_PROT_EXECUTE | VM_PROT_READ;
1646
1647         /*
1648          * If FWRITE's set, we can allow VM_PROT_WRITE unless it's a shared
1649          * mapping with a write seal applied.  Private mappings are always
1650          * writeable.
1651          */
1652         if ((flags & MAP_SHARED) == 0) {
1653                 cap_maxprot |= VM_PROT_WRITE;
1654                 maxprot |= VM_PROT_WRITE;
1655                 writecnt = false;
1656         } else {
1657                 if ((fp->f_flag & FWRITE) != 0 &&
1658                     (shmfd->shm_seals & F_SEAL_WRITE) == 0)
1659                         maxprot |= VM_PROT_WRITE;
1660
1661                 /*
1662                  * Any mappings from a writable descriptor may be upgraded to
1663                  * VM_PROT_WRITE with mprotect(2), unless a write-seal was
1664                  * applied between the open and subsequent mmap(2).  We want to
1665                  * reject application of a write seal as long as any such
1666                  * mapping exists so that the seal cannot be trivially bypassed.
1667                  */
1668                 writecnt = (maxprot & VM_PROT_WRITE) != 0;
1669                 if (!writecnt && (prot & VM_PROT_WRITE) != 0) {
1670                         error = EACCES;
1671                         goto out;
1672                 }
1673         }
1674         maxprot &= cap_maxprot;
1675
1676         /* See comment in vn_mmap(). */
1677         if (
1678 #ifdef _LP64
1679             objsize > OFF_MAX ||
1680 #endif
1681             foff > OFF_MAX - objsize) {
1682                 error = EINVAL;
1683                 goto out;
1684         }
1685
1686 #ifdef MAC
1687         error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags);
1688         if (error != 0)
1689                 goto out;
1690 #endif
1691
1692         mtx_lock(&shm_timestamp_lock);
1693         vfs_timestamp(&shmfd->shm_atime);
1694         mtx_unlock(&shm_timestamp_lock);
1695         vm_object_reference(shmfd->shm_object);
1696
1697         if (shm_largepage(shmfd)) {
1698                 writecnt = false;
1699                 error = shm_mmap_large(shmfd, map, addr, objsize, prot,
1700                     maxprot, flags, foff, td);
1701         } else {
1702                 if (writecnt) {
1703                         vm_pager_update_writecount(shmfd->shm_object, 0,
1704                             objsize);
1705                 }
1706                 error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags,
1707                     shmfd->shm_object, foff, writecnt, td);
1708         }
1709         if (error != 0) {
1710                 if (writecnt)
1711                         vm_pager_release_writecount(shmfd->shm_object, 0,
1712                             objsize);
1713                 vm_object_deallocate(shmfd->shm_object);
1714         }
1715 out:
1716         rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
1717         return (error);
1718 }
1719
1720 static int
1721 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
1722     struct thread *td)
1723 {
1724         struct shmfd *shmfd;
1725         int error;
1726
1727         error = 0;
1728         shmfd = fp->f_data;
1729         mtx_lock(&shm_timestamp_lock);
1730         /*
1731          * SUSv4 says that x bits of permission need not be affected.
1732          * Be consistent with our shm_open there.
1733          */
1734 #ifdef MAC
1735         error = mac_posixshm_check_setmode(active_cred, shmfd, mode);
1736         if (error != 0)
1737                 goto out;
1738 #endif
1739         error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
1740             VADMIN, active_cred);
1741         if (error != 0)
1742                 goto out;
1743         shmfd->shm_mode = mode & ACCESSPERMS;
1744 out:
1745         mtx_unlock(&shm_timestamp_lock);
1746         return (error);
1747 }
1748
1749 static int
1750 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
1751     struct thread *td)
1752 {
1753         struct shmfd *shmfd;
1754         int error;
1755
1756         error = 0;
1757         shmfd = fp->f_data;
1758         mtx_lock(&shm_timestamp_lock);
1759 #ifdef MAC
1760         error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid);
1761         if (error != 0)
1762                 goto out;
1763 #endif
1764         if (uid == (uid_t)-1)
1765                 uid = shmfd->shm_uid;
1766         if (gid == (gid_t)-1)
1767                  gid = shmfd->shm_gid;
1768         if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) ||
1769             (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) &&
1770             (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN)))
1771                 goto out;
1772         shmfd->shm_uid = uid;
1773         shmfd->shm_gid = gid;
1774 out:
1775         mtx_unlock(&shm_timestamp_lock);
1776         return (error);
1777 }
1778
1779 /*
1780  * Helper routines to allow the backing object of a shared memory file
1781  * descriptor to be mapped in the kernel.
1782  */
1783 int
1784 shm_map(struct file *fp, size_t size, off_t offset, void **memp)
1785 {
1786         struct shmfd *shmfd;
1787         vm_offset_t kva, ofs;
1788         vm_object_t obj;
1789         int rv;
1790
1791         if (fp->f_type != DTYPE_SHM)
1792                 return (EINVAL);
1793         shmfd = fp->f_data;
1794         obj = shmfd->shm_object;
1795         VM_OBJECT_WLOCK(obj);
1796         /*
1797          * XXXRW: This validation is probably insufficient, and subject to
1798          * sign errors.  It should be fixed.
1799          */
1800         if (offset >= shmfd->shm_size ||
1801             offset + size > round_page(shmfd->shm_size)) {
1802                 VM_OBJECT_WUNLOCK(obj);
1803                 return (EINVAL);
1804         }
1805
1806         shmfd->shm_kmappings++;
1807         vm_object_reference_locked(obj);
1808         VM_OBJECT_WUNLOCK(obj);
1809
1810         /* Map the object into the kernel_map and wire it. */
1811         kva = vm_map_min(kernel_map);
1812         ofs = offset & PAGE_MASK;
1813         offset = trunc_page(offset);
1814         size = round_page(size + ofs);
1815         rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0,
1816             VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE,
1817             VM_PROT_READ | VM_PROT_WRITE, 0);
1818         if (rv == KERN_SUCCESS) {
1819                 rv = vm_map_wire(kernel_map, kva, kva + size,
1820                     VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
1821                 if (rv == KERN_SUCCESS) {
1822                         *memp = (void *)(kva + ofs);
1823                         return (0);
1824                 }
1825                 vm_map_remove(kernel_map, kva, kva + size);
1826         } else
1827                 vm_object_deallocate(obj);
1828
1829         /* On failure, drop our mapping reference. */
1830         VM_OBJECT_WLOCK(obj);
1831         shmfd->shm_kmappings--;
1832         VM_OBJECT_WUNLOCK(obj);
1833
1834         return (vm_mmap_to_errno(rv));
1835 }
1836
1837 /*
1838  * We require the caller to unmap the entire entry.  This allows us to
1839  * safely decrement shm_kmappings when a mapping is removed.
1840  */
1841 int
1842 shm_unmap(struct file *fp, void *mem, size_t size)
1843 {
1844         struct shmfd *shmfd;
1845         vm_map_entry_t entry;
1846         vm_offset_t kva, ofs;
1847         vm_object_t obj;
1848         vm_pindex_t pindex;
1849         vm_prot_t prot;
1850         boolean_t wired;
1851         vm_map_t map;
1852         int rv;
1853
1854         if (fp->f_type != DTYPE_SHM)
1855                 return (EINVAL);
1856         shmfd = fp->f_data;
1857         kva = (vm_offset_t)mem;
1858         ofs = kva & PAGE_MASK;
1859         kva = trunc_page(kva);
1860         size = round_page(size + ofs);
1861         map = kernel_map;
1862         rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry,
1863             &obj, &pindex, &prot, &wired);
1864         if (rv != KERN_SUCCESS)
1865                 return (EINVAL);
1866         if (entry->start != kva || entry->end != kva + size) {
1867                 vm_map_lookup_done(map, entry);
1868                 return (EINVAL);
1869         }
1870         vm_map_lookup_done(map, entry);
1871         if (obj != shmfd->shm_object)
1872                 return (EINVAL);
1873         vm_map_remove(map, kva, kva + size);
1874         VM_OBJECT_WLOCK(obj);
1875         KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped"));
1876         shmfd->shm_kmappings--;
1877         VM_OBJECT_WUNLOCK(obj);
1878         return (0);
1879 }
1880
1881 static int
1882 shm_fill_kinfo_locked(struct shmfd *shmfd, struct kinfo_file *kif, bool list)
1883 {
1884         const char *path, *pr_path;
1885         size_t pr_pathlen;
1886         bool visible;
1887
1888         sx_assert(&shm_dict_lock, SA_LOCKED);
1889         kif->kf_type = KF_TYPE_SHM;
1890         kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode;
1891         kif->kf_un.kf_file.kf_file_size = shmfd->shm_size;
1892         if (shmfd->shm_path != NULL) {
1893                 if (shmfd->shm_path != NULL) {
1894                         path = shmfd->shm_path;
1895                         pr_path = curthread->td_ucred->cr_prison->pr_path;
1896                         if (strcmp(pr_path, "/") != 0) {
1897                                 /* Return the jail-rooted pathname. */
1898                                 pr_pathlen = strlen(pr_path);
1899                                 visible = strncmp(path, pr_path, pr_pathlen)
1900                                     == 0 && path[pr_pathlen] == '/';
1901                                 if (list && !visible)
1902                                         return (EPERM);
1903                                 if (visible)
1904                                         path += pr_pathlen;
1905                         }
1906                         strlcpy(kif->kf_path, path, sizeof(kif->kf_path));
1907                 }
1908         }
1909         return (0);
1910 }
1911
1912 static int
1913 shm_fill_kinfo(struct file *fp, struct kinfo_file *kif,
1914     struct filedesc *fdp __unused)
1915 {
1916         int res;
1917
1918         sx_slock(&shm_dict_lock);
1919         res = shm_fill_kinfo_locked(fp->f_data, kif, false);
1920         sx_sunlock(&shm_dict_lock);
1921         return (res);
1922 }
1923
1924 static int
1925 shm_add_seals(struct file *fp, int seals)
1926 {
1927         struct shmfd *shmfd;
1928         void *rl_cookie;
1929         vm_ooffset_t writemappings;
1930         int error, nseals;
1931
1932         error = 0;
1933         shmfd = fp->f_data;
1934         rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
1935             &shmfd->shm_mtx);
1936
1937         /* Even already-set seals should result in EPERM. */
1938         if ((shmfd->shm_seals & F_SEAL_SEAL) != 0) {
1939                 error = EPERM;
1940                 goto out;
1941         }
1942         nseals = seals & ~shmfd->shm_seals;
1943         if ((nseals & F_SEAL_WRITE) != 0) {
1944                 if (shm_largepage(shmfd)) {
1945                         error = ENOTSUP;
1946                         goto out;
1947                 }
1948
1949                 /*
1950                  * The rangelock above prevents writable mappings from being
1951                  * added after we've started applying seals.  The RLOCK here
1952                  * is to avoid torn reads on ILP32 arches as unmapping/reducing
1953                  * writemappings will be done without a rangelock.
1954                  */
1955                 VM_OBJECT_RLOCK(shmfd->shm_object);
1956                 writemappings = shmfd->shm_object->un_pager.swp.writemappings;
1957                 VM_OBJECT_RUNLOCK(shmfd->shm_object);
1958                 /* kmappings are also writable */
1959                 if (writemappings > 0) {
1960                         error = EBUSY;
1961                         goto out;
1962                 }
1963         }
1964         shmfd->shm_seals |= nseals;
1965 out:
1966         rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
1967         return (error);
1968 }
1969
1970 static int
1971 shm_get_seals(struct file *fp, int *seals)
1972 {
1973         struct shmfd *shmfd;
1974
1975         shmfd = fp->f_data;
1976         *seals = shmfd->shm_seals;
1977         return (0);
1978 }
1979
1980 static int
1981 shm_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
1982 {
1983         void *rl_cookie;
1984         struct shmfd *shmfd;
1985         size_t size;
1986         int error;
1987
1988         /* This assumes that the caller already checked for overflow. */
1989         error = 0;
1990         shmfd = fp->f_data;
1991         size = offset + len;
1992
1993         /*
1994          * Just grab the rangelock for the range that we may be attempting to
1995          * grow, rather than blocking read/write for regions we won't be
1996          * touching while this (potential) resize is in progress.  Other
1997          * attempts to resize the shmfd will have to take a write lock from 0 to
1998          * OFF_MAX, so this being potentially beyond the current usable range of
1999          * the shmfd is not necessarily a concern.  If other mechanisms are
2000          * added to grow a shmfd, this may need to be re-evaluated.
2001          */
2002         rl_cookie = rangelock_wlock(&shmfd->shm_rl, offset, size,
2003             &shmfd->shm_mtx);
2004         if (size > shmfd->shm_size)
2005                 error = shm_dotruncate_cookie(shmfd, size, rl_cookie);
2006         rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
2007         /* Translate to posix_fallocate(2) return value as needed. */
2008         if (error == ENOMEM)
2009                 error = ENOSPC;
2010         return (error);
2011 }
2012
2013 static int
2014 sysctl_posix_shm_list(SYSCTL_HANDLER_ARGS)
2015 {
2016         struct shm_mapping *shmm;
2017         struct sbuf sb;
2018         struct kinfo_file kif;
2019         u_long i;
2020         ssize_t curlen;
2021         int error, error2;
2022
2023         sbuf_new_for_sysctl(&sb, NULL, sizeof(struct kinfo_file) * 5, req);
2024         sbuf_clear_flags(&sb, SBUF_INCLUDENUL);
2025         curlen = 0;
2026         error = 0;
2027         sx_slock(&shm_dict_lock);
2028         for (i = 0; i < shm_hash + 1; i++) {
2029                 LIST_FOREACH(shmm, &shm_dictionary[i], sm_link) {
2030                         error = shm_fill_kinfo_locked(shmm->sm_shmfd,
2031                             &kif, true);
2032                         if (error == EPERM) {
2033                                 error = 0;
2034                                 continue;
2035                         }
2036                         if (error != 0)
2037                                 break;
2038                         pack_kinfo(&kif);
2039                         error = sbuf_bcat(&sb, &kif, kif.kf_structsize) == 0 ?
2040                             0 : ENOMEM;
2041                         if (error != 0)
2042                                 break;
2043                         curlen += kif.kf_structsize;
2044                 }
2045         }
2046         sx_sunlock(&shm_dict_lock);
2047         error2 = sbuf_finish(&sb);
2048         sbuf_delete(&sb);
2049         return (error != 0 ? error : error2);
2050 }
2051
2052 SYSCTL_PROC(_kern_ipc, OID_AUTO, posix_shm_list,
2053     CTLFLAG_RD | CTLFLAG_PRISON | CTLFLAG_MPSAFE | CTLTYPE_OPAQUE,
2054     NULL, 0, sysctl_posix_shm_list, "",
2055     "POSIX SHM list");
2056
2057 int
2058 kern_shm_open(struct thread *td, const char *path, int flags, mode_t mode,
2059     struct filecaps *caps)
2060 {
2061
2062         return (kern_shm_open2(td, path, flags, mode, 0, caps, NULL));
2063 }
2064
2065 /*
2066  * This version of the shm_open() interface leaves CLOEXEC behavior up to the
2067  * caller, and libc will enforce it for the traditional shm_open() call.  This
2068  * allows other consumers, like memfd_create(), to opt-in for CLOEXEC.  This
2069  * interface also includes a 'name' argument that is currently unused, but could
2070  * potentially be exported later via some interface for debugging purposes.
2071  * From the kernel's perspective, it is optional.  Individual consumers like
2072  * memfd_create() may require it in order to be compatible with other systems
2073  * implementing the same function.
2074  */
2075 int
2076 sys_shm_open2(struct thread *td, struct shm_open2_args *uap)
2077 {
2078
2079         return (kern_shm_open2(td, uap->path, uap->flags, uap->mode,
2080             uap->shmflags, NULL, uap->name));
2081 }