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