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 __diagused;
 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 const 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;
 737         vm_pindex_t oldobjsz __unused;
 738         int aflags, error, i, psind, try;
 739
 740         KASSERT(length >= 0, ("shm_dotruncate: length < 0"));
 741         object = shmfd->shm_object;
 742         VM_OBJECT_ASSERT_WLOCKED(object);
 743         rangelock_cookie_assert(rl_cookie, RA_WLOCKED);
 744
 745         oldobjsz = object->size;
 746         newobjsz = OFF_TO_IDX(length);
 747         if (length == shmfd->shm_size)
 748                 return (0);
 749         psind = shmfd->shm_lp_psind;
 750         if (psind == 0 && length != 0)
 751                 return (EINVAL);
 752         if ((length & (pagesizes[psind] - 1)) != 0)
 753                 return (EINVAL);
 754
 755         if (length < shmfd->shm_size) {
 756                 if ((shmfd->shm_seals & F_SEAL_SHRINK) != 0)
 757                         return (EPERM);
 758                 if (shmfd->shm_kmappings > 0)
 759                         return (EBUSY);
 760                 return (ENOTSUP);       /* Pages are unmanaged. */
 761 #if 0
 762                 vm_object_page_remove(object, newobjsz, oldobjsz, 0);
 763                 object->size = newobjsz;
 764                 shmfd->shm_size = length;
 765                 return (0);
 766 #endif
 767         }
 768
 769         if ((shmfd->shm_seals & F_SEAL_GROW) != 0)
 770                 return (EPERM);
 771
 772         aflags = VM_ALLOC_NORMAL | VM_ALLOC_ZERO;
 773         if (shmfd->shm_lp_alloc_policy == SHM_LARGEPAGE_ALLOC_NOWAIT)
 774                 aflags |= VM_ALLOC_WAITFAIL;
 775         try = 0;
 776
 777         /*
 778          * Extend shmfd and object, keeping all already fully
 779          * allocated large pages intact even on error, because dropped
 780          * object lock might allowed mapping of them.
 781          */
 782         while (object->size < newobjsz) {
 783                 m = vm_page_alloc_contig(object, object->size, aflags,
 784                     pagesizes[psind] / PAGE_SIZE, 0, ~0,
 785                     pagesizes[psind], 0,
 786                     VM_MEMATTR_DEFAULT);
 787                 if (m == NULL) {
 788                         VM_OBJECT_WUNLOCK(object);
 789                         if (shmfd->shm_lp_alloc_policy ==
 790                             SHM_LARGEPAGE_ALLOC_NOWAIT ||
 791                             (shmfd->shm_lp_alloc_policy ==
 792                             SHM_LARGEPAGE_ALLOC_DEFAULT &&
 793                             try >= largepage_reclaim_tries)) {
 794                                 VM_OBJECT_WLOCK(object);
 795                                 return (ENOMEM);
 796                         }
 797                         error = vm_page_reclaim_contig(aflags,
 798                             pagesizes[psind] / PAGE_SIZE, 0, ~0,
 799                             pagesizes[psind], 0) ? 0 :
 800                             vm_wait_intr(object);
 801                         if (error != 0) {
 802                                 VM_OBJECT_WLOCK(object);
 803                                 return (error);
 804                         }
 805                         try++;
 806                         VM_OBJECT_WLOCK(object);
 807                         continue;
 808                 }
 809                 try = 0;
 810                 for (i = 0; i < pagesizes[psind] / PAGE_SIZE; i++) {
 811                         if ((m[i].flags & PG_ZERO) == 0)
 812                                 pmap_zero_page(&m[i]);
 813                         vm_page_valid(&m[i]);
 814                         vm_page_xunbusy(&m[i]);
 815                 }
 816                 object->size += OFF_TO_IDX(pagesizes[psind]);
 817                 shmfd->shm_size += pagesizes[psind];
 818                 atomic_add_long(&count_largepages[psind], 1);
 819                 vm_wire_add(atop(pagesizes[psind]));
 820         }
 821         return (0);
 822 }
 823
 824 static int
 825 shm_dotruncate_cookie(struct shmfd *shmfd, off_t length, void *rl_cookie)
 826 {
 827         int error;
 828
 829         VM_OBJECT_WLOCK(shmfd->shm_object);
 830         error = shm_largepage(shmfd) ? shm_dotruncate_largepage(shmfd,
 831             length, rl_cookie) : shm_dotruncate_locked(shmfd, length,
 832             rl_cookie);
 833         VM_OBJECT_WUNLOCK(shmfd->shm_object);
 834         return (error);
 835 }
 836
 837 int
 838 shm_dotruncate(struct shmfd *shmfd, off_t length)
 839 {
 840         void *rl_cookie;
 841         int error;
 842
 843         rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
 844             &shmfd->shm_mtx);
 845         error = shm_dotruncate_cookie(shmfd, length, rl_cookie);
 846         rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
 847         return (error);
 848 }
 849
 850 /*
 851  * shmfd object management including creation and reference counting
 852  * routines.
 853  */
 854 struct shmfd *
 855 shm_alloc(struct ucred *ucred, mode_t mode, bool largepage)
 856 {
 857         struct shmfd *shmfd;
 858
 859         shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO);
 860         shmfd->shm_size = 0;
 861         shmfd->shm_uid = ucred->cr_uid;
 862         shmfd->shm_gid = ucred->cr_gid;
 863         shmfd->shm_mode = mode;
 864         if (largepage) {
 865                 shmfd->shm_object = phys_pager_allocate(NULL,
 866                     &shm_largepage_phys_ops, NULL, shmfd->shm_size,
 867                     VM_PROT_DEFAULT, 0, ucred);
 868                 shmfd->shm_lp_alloc_policy = SHM_LARGEPAGE_ALLOC_DEFAULT;
 869         } else {
 870                 shmfd->shm_object = vm_pager_allocate(OBJT_SWAP, NULL,
 871                     shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred);
 872         }
 873         KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate"));
 874         vfs_timestamp(&shmfd->shm_birthtime);
 875         shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime =
 876             shmfd->shm_birthtime;
 877         shmfd->shm_ino = alloc_unr64(&shm_ino_unr);
 878         refcount_init(&shmfd->shm_refs, 1);
 879         mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF);
 880         rangelock_init(&shmfd->shm_rl);
 881 #ifdef MAC
 882         mac_posixshm_init(shmfd);
 883         mac_posixshm_create(ucred, shmfd);
 884 #endif
 885
 886         return (shmfd);
 887 }
 888
 889 struct shmfd *
 890 shm_hold(struct shmfd *shmfd)
 891 {
 892
 893         refcount_acquire(&shmfd->shm_refs);
 894         return (shmfd);
 895 }
 896
 897 void
 898 shm_drop(struct shmfd *shmfd)
 899 {
 900
 901         if (refcount_release(&shmfd->shm_refs)) {
 902 #ifdef MAC
 903                 mac_posixshm_destroy(shmfd);
 904 #endif
 905                 rangelock_destroy(&shmfd->shm_rl);
 906                 mtx_destroy(&shmfd->shm_mtx);
 907                 vm_object_deallocate(shmfd->shm_object);
 908                 free(shmfd, M_SHMFD);
 909         }
 910 }
 911
 912 /*
 913  * Determine if the credentials have sufficient permissions for a
 914  * specified combination of FREAD and FWRITE.
 915  */
 916 int
 917 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags)
 918 {
 919         accmode_t accmode;
 920         int error;
 921
 922         accmode = 0;
 923         if (flags & FREAD)
 924                 accmode |= VREAD;
 925         if (flags & FWRITE)
 926                 accmode |= VWRITE;
 927         mtx_lock(&shm_timestamp_lock);
 928         error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
 929             accmode, ucred);
 930         mtx_unlock(&shm_timestamp_lock);
 931         return (error);
 932 }
 933
 934 static void
 935 shm_init(void *arg)
 936 {
 937         char name[32];
 938         int i;
 939
 940         mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF);
 941         sx_init(&shm_dict_lock, "shm dictionary");
 942         shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash);
 943         new_unrhdr64(&shm_ino_unr, 1);
 944         shm_dev_ino = devfs_alloc_cdp_inode();
 945         KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized"));
 946
 947         for (i = 1; i < MAXPAGESIZES; i++) {
 948                 if (pagesizes[i] == 0)
 949                         break;
 950 #define M       (1024 * 1024)
 951 #define G       (1024 * M)
 952                 if (pagesizes[i] >= G)
 953                         snprintf(name, sizeof(name), "%luG", pagesizes[i] / G);
 954                 else if (pagesizes[i] >= M)
 955                         snprintf(name, sizeof(name), "%luM", pagesizes[i] / M);
 956                 else
 957                         snprintf(name, sizeof(name), "%lu", pagesizes[i]);
 958 #undef G
 959 #undef M
 960                 SYSCTL_ADD_ULONG(NULL, SYSCTL_STATIC_CHILDREN(_vm_largepages),
 961                     OID_AUTO, name, CTLFLAG_RD, &count_largepages[i],
 962                     "number of non-transient largepages allocated");
 963         }
 964 }
 965 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL);
 966
 967 /*
 968  * Dictionary management.  We maintain an in-kernel dictionary to map
 969  * paths to shmfd objects.  We use the FNV hash on the path to store
 970  * the mappings in a hash table.
 971  */
 972 static struct shmfd *
 973 shm_lookup(char *path, Fnv32_t fnv)
 974 {
 975         struct shm_mapping *map;
 976
 977         LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
 978                 if (map->sm_fnv != fnv)
 979                         continue;
 980                 if (strcmp(map->sm_path, path) == 0)
 981                         return (map->sm_shmfd);
 982         }
 983
 984         return (NULL);
 985 }
 986
 987 static void
 988 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd)
 989 {
 990         struct shm_mapping *map;
 991
 992         map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK);
 993         map->sm_path = path;
 994         map->sm_fnv = fnv;
 995         map->sm_shmfd = shm_hold(shmfd);
 996         shmfd->shm_path = path;
 997         LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link);
 998 }
 999
1000 static int
1001 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred)
1002 {
1003         struct shm_mapping *map;
1004         int error;
1005
1006         LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
1007                 if (map->sm_fnv != fnv)
1008                         continue;
1009                 if (strcmp(map->sm_path, path) == 0) {
1010 #ifdef MAC
1011                         error = mac_posixshm_check_unlink(ucred, map->sm_shmfd);
1012                         if (error)
1013                                 return (error);
1014 #endif
1015                         error = shm_access(map->sm_shmfd, ucred,
1016                             FREAD | FWRITE);
1017                         if (error)
1018                                 return (error);
1019                         map->sm_shmfd->shm_path = NULL;
1020                         LIST_REMOVE(map, sm_link);
1021                         shm_drop(map->sm_shmfd);
1022                         free(map->sm_path, M_SHMFD);
1023                         free(map, M_SHMFD);
1024                         return (0);
1025                 }
1026         }
1027
1028         return (ENOENT);
1029 }
1030
1031 int
1032 kern_shm_open2(struct thread *td, const char *userpath, int flags, mode_t mode,
1033     int shmflags, struct filecaps *fcaps, const char *name __unused)
1034 {
1035         struct pwddesc *pdp;
1036         struct shmfd *shmfd;
1037         struct file *fp;
1038         char *path;
1039         void *rl_cookie;
1040         Fnv32_t fnv;
1041         mode_t cmode;
1042         int error, fd, initial_seals;
1043         bool largepage;
1044
1045         if ((shmflags & ~(SHM_ALLOW_SEALING | SHM_GROW_ON_WRITE |
1046             SHM_LARGEPAGE)) != 0)
1047                 return (EINVAL);
1048
1049         initial_seals = F_SEAL_SEAL;
1050         if ((shmflags & SHM_ALLOW_SEALING) != 0)
1051                 initial_seals &= ~F_SEAL_SEAL;
1052
1053 #ifdef CAPABILITY_MODE
1054         /*
1055          * shm_open(2) is only allowed for anonymous objects.
1056          */
1057         if (IN_CAPABILITY_MODE(td) && (userpath != SHM_ANON))
1058                 return (ECAPMODE);
1059 #endif
1060
1061         AUDIT_ARG_FFLAGS(flags);
1062         AUDIT_ARG_MODE(mode);
1063
1064         if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR)
1065                 return (EINVAL);
1066
1067         if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0)
1068                 return (EINVAL);
1069
1070         largepage = (shmflags & SHM_LARGEPAGE) != 0;
1071         if (largepage && !PMAP_HAS_LARGEPAGES)
1072                 return (ENOTTY);
1073
1074         /*
1075          * Currently only F_SEAL_SEAL may be set when creating or opening shmfd.
1076          * If the decision is made later to allow additional seals, care must be
1077          * taken below to ensure that the seals are properly set if the shmfd
1078          * already existed -- this currently assumes that only F_SEAL_SEAL can
1079          * be set and doesn't take further precautions to ensure the validity of
1080          * the seals being added with respect to current mappings.
1081          */
1082         if ((initial_seals & ~F_SEAL_SEAL) != 0)
1083                 return (EINVAL);
1084
1085         pdp = td->td_proc->p_pd;
1086         cmode = (mode & ~pdp->pd_cmask) & ACCESSPERMS;
1087
1088         /*
1089          * shm_open(2) created shm should always have O_CLOEXEC set, as mandated
1090          * by POSIX.  We allow it to be unset here so that an in-kernel
1091          * interface may be written as a thin layer around shm, optionally not
1092          * setting CLOEXEC.  For shm_open(2), O_CLOEXEC is set unconditionally
1093          * in sys_shm_open() to keep this implementation compliant.
1094          */
1095         error = falloc_caps(td, &fp, &fd, flags & O_CLOEXEC, fcaps);
1096         if (error)
1097                 return (error);
1098
1099         /* A SHM_ANON path pointer creates an anonymous object. */
1100         if (userpath == SHM_ANON) {
1101                 /* A read-only anonymous object is pointless. */
1102                 if ((flags & O_ACCMODE) == O_RDONLY) {
1103                         fdclose(td, fp, fd);
1104                         fdrop(fp, td);
1105                         return (EINVAL);
1106                 }
1107                 shmfd = shm_alloc(td->td_ucred, cmode, largepage);
1108                 shmfd->shm_seals = initial_seals;
1109                 shmfd->shm_flags = shmflags;
1110         } else {
1111                 error = shm_copyin_path(td, userpath, &path);
1112                 if (error != 0) {
1113                         fdclose(td, fp, fd);
1114                         fdrop(fp, td);
1115                         return (error);
1116                 }
1117
1118                 AUDIT_ARG_UPATH1_CANON(path);
1119                 fnv = fnv_32_str(path, FNV1_32_INIT);
1120                 sx_xlock(&shm_dict_lock);
1121                 shmfd = shm_lookup(path, fnv);
1122                 if (shmfd == NULL) {
1123                         /* Object does not yet exist, create it if requested. */
1124                         if (flags & O_CREAT) {
1125 #ifdef MAC
1126                                 error = mac_posixshm_check_create(td->td_ucred,
1127                                     path);
1128                                 if (error == 0) {
1129 #endif
1130                                         shmfd = shm_alloc(td->td_ucred, cmode,
1131                                             largepage);
1132                                         shmfd->shm_seals = initial_seals;
1133                                         shmfd->shm_flags = shmflags;
1134                                         shm_insert(path, fnv, shmfd);
1135 #ifdef MAC
1136                                 }
1137 #endif
1138                         } else {
1139                                 free(path, M_SHMFD);
1140                                 error = ENOENT;
1141                         }
1142                 } else {
1143                         rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
1144                             &shmfd->shm_mtx);
1145
1146                         /*
1147                          * kern_shm_open() likely shouldn't ever error out on
1148                          * trying to set a seal that already exists, unlike
1149                          * F_ADD_SEALS.  This would break terribly as
1150                          * shm_open(2) actually sets F_SEAL_SEAL to maintain
1151                          * historical behavior where the underlying file could
1152                          * not be sealed.
1153                          */
1154                         initial_seals &= ~shmfd->shm_seals;
1155
1156                         /*
1157                          * Object already exists, obtain a new
1158                          * reference if requested and permitted.
1159                          */
1160                         free(path, M_SHMFD);
1161
1162                         /*
1163                          * initial_seals can't set additional seals if we've
1164                          * already been set F_SEAL_SEAL.  If F_SEAL_SEAL is set,
1165                          * then we've already removed that one from
1166                          * initial_seals.  This is currently redundant as we
1167                          * only allow setting F_SEAL_SEAL at creation time, but
1168                          * it's cheap to check and decreases the effort required
1169                          * to allow additional seals.
1170                          */
1171                         if ((shmfd->shm_seals & F_SEAL_SEAL) != 0 &&
1172                             initial_seals != 0)
1173                                 error = EPERM;
1174                         else if ((flags & (O_CREAT | O_EXCL)) ==
1175                             (O_CREAT | O_EXCL))
1176                                 error = EEXIST;
1177                         else if (shmflags != 0 && shmflags != shmfd->shm_flags)
1178                                 error = EINVAL;
1179                         else {
1180 #ifdef MAC
1181                                 error = mac_posixshm_check_open(td->td_ucred,
1182                                     shmfd, FFLAGS(flags & O_ACCMODE));
1183                                 if (error == 0)
1184 #endif
1185                                 error = shm_access(shmfd, td->td_ucred,
1186                                     FFLAGS(flags & O_ACCMODE));
1187                         }
1188
1189                         /*
1190                          * Truncate the file back to zero length if
1191                          * O_TRUNC was specified and the object was
1192                          * opened with read/write.
1193                          */
1194                         if (error == 0 &&
1195                             (flags & (O_ACCMODE | O_TRUNC)) ==
1196                             (O_RDWR | O_TRUNC)) {
1197                                 VM_OBJECT_WLOCK(shmfd->shm_object);
1198 #ifdef MAC
1199                                 error = mac_posixshm_check_truncate(
1200                                         td->td_ucred, fp->f_cred, shmfd);
1201                                 if (error == 0)
1202 #endif
1203                                         error = shm_dotruncate_locked(shmfd, 0,
1204                                             rl_cookie);
1205                                 VM_OBJECT_WUNLOCK(shmfd->shm_object);
1206                         }
1207                         if (error == 0) {
1208                                 /*
1209                                  * Currently we only allow F_SEAL_SEAL to be
1210                                  * set initially.  As noted above, this would
1211                                  * need to be reworked should that change.
1212                                  */
1213                                 shmfd->shm_seals |= initial_seals;
1214                                 shm_hold(shmfd);
1215                         }
1216                         rangelock_unlock(&shmfd->shm_rl, rl_cookie,
1217                             &shmfd->shm_mtx);
1218                 }
1219                 sx_xunlock(&shm_dict_lock);
1220
1221                 if (error) {
1222                         fdclose(td, fp, fd);
1223                         fdrop(fp, td);
1224                         return (error);
1225                 }
1226         }
1227
1228         finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops);
1229
1230         td->td_retval[0] = fd;
1231         fdrop(fp, td);
1232
1233         return (0);
1234 }
1235
1236 /* System calls. */
1237 #ifdef COMPAT_FREEBSD12
1238 int
1239 freebsd12_shm_open(struct thread *td, struct freebsd12_shm_open_args *uap)
1240 {
1241
1242         return (kern_shm_open(td, uap->path, uap->flags | O_CLOEXEC,
1243             uap->mode, NULL));
1244 }
1245 #endif
1246
1247 int
1248 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap)
1249 {
1250         char *path;
1251         Fnv32_t fnv;
1252         int error;
1253
1254         error = shm_copyin_path(td, uap->path, &path);
1255         if (error != 0)
1256                 return (error);
1257
1258         AUDIT_ARG_UPATH1_CANON(path);
1259         fnv = fnv_32_str(path, FNV1_32_INIT);
1260         sx_xlock(&shm_dict_lock);
1261         error = shm_remove(path, fnv, td->td_ucred);
1262         sx_xunlock(&shm_dict_lock);
1263         free(path, M_SHMFD);
1264
1265         return (error);
1266 }
1267
1268 int
1269 sys_shm_rename(struct thread *td, struct shm_rename_args *uap)
1270 {
1271         char *path_from = NULL, *path_to = NULL;
1272         Fnv32_t fnv_from, fnv_to;
1273         struct shmfd *fd_from;
1274         struct shmfd *fd_to;
1275         int error;
1276         int flags;
1277
1278         flags = uap->flags;
1279         AUDIT_ARG_FFLAGS(flags);
1280
1281         /*
1282          * Make sure the user passed only valid flags.
1283          * If you add a new flag, please add a new term here.
1284          */
1285         if ((flags & ~(
1286             SHM_RENAME_NOREPLACE |
1287             SHM_RENAME_EXCHANGE
1288             )) != 0) {
1289                 error = EINVAL;
1290                 goto out;
1291         }
1292
1293         /*
1294          * EXCHANGE and NOREPLACE don't quite make sense together. Let's
1295          * force the user to choose one or the other.
1296          */
1297         if ((flags & SHM_RENAME_NOREPLACE) != 0 &&
1298             (flags & SHM_RENAME_EXCHANGE) != 0) {
1299                 error = EINVAL;
1300                 goto out;
1301         }
1302
1303         /* Renaming to or from anonymous makes no sense */
1304         if (uap->path_from == SHM_ANON || uap->path_to == SHM_ANON) {
1305                 error = EINVAL;
1306                 goto out;
1307         }
1308
1309         error = shm_copyin_path(td, uap->path_from, &path_from);
1310         if (error != 0)
1311                 goto out;
1312
1313         error = shm_copyin_path(td, uap->path_to, &path_to);
1314         if (error != 0)
1315                 goto out;
1316
1317         AUDIT_ARG_UPATH1_CANON(path_from);
1318         AUDIT_ARG_UPATH2_CANON(path_to);
1319
1320         /* Rename with from/to equal is a no-op */
1321         if (strcmp(path_from, path_to) == 0)
1322                 goto out;
1323
1324         fnv_from = fnv_32_str(path_from, FNV1_32_INIT);
1325         fnv_to = fnv_32_str(path_to, FNV1_32_INIT);
1326
1327         sx_xlock(&shm_dict_lock);
1328
1329         fd_from = shm_lookup(path_from, fnv_from);
1330         if (fd_from == NULL) {
1331                 error = ENOENT;
1332                 goto out_locked;
1333         }
1334
1335         fd_to = shm_lookup(path_to, fnv_to);
1336         if ((flags & SHM_RENAME_NOREPLACE) != 0 && fd_to != NULL) {
1337                 error = EEXIST;
1338                 goto out_locked;
1339         }
1340
1341         /*
1342          * Unconditionally prevents shm_remove from invalidating the 'from'
1343          * shm's state.
1344          */
1345         shm_hold(fd_from);
1346         error = shm_remove(path_from, fnv_from, td->td_ucred);
1347
1348         /*
1349          * One of my assumptions failed if ENOENT (e.g. locking didn't
1350          * protect us)
1351          */
1352         KASSERT(error != ENOENT, ("Our shm disappeared during shm_rename: %s",
1353             path_from));
1354         if (error != 0) {
1355                 shm_drop(fd_from);
1356                 goto out_locked;
1357         }
1358
1359         /*
1360          * If we are exchanging, we need to ensure the shm_remove below
1361          * doesn't invalidate the dest shm's state.
1362          */
1363         if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL)
1364                 shm_hold(fd_to);
1365
1366         /*
1367          * NOTE: if path_to is not already in the hash, c'est la vie;
1368          * it simply means we have nothing already at path_to to unlink.
1369          * That is the ENOENT case.
1370          *
1371          * If we somehow don't have access to unlink this guy, but
1372          * did for the shm at path_from, then relink the shm to path_from
1373          * and abort with EACCES.
1374          *
1375          * All other errors: that is weird; let's relink and abort the
1376          * operation.
1377          */
1378         error = shm_remove(path_to, fnv_to, td->td_ucred);
1379         if (error != 0 && error != ENOENT) {
1380                 shm_insert(path_from, fnv_from, fd_from);
1381                 shm_drop(fd_from);
1382                 /* Don't free path_from now, since the hash references it */
1383                 path_from = NULL;
1384                 goto out_locked;
1385         }
1386
1387         error = 0;
1388
1389         shm_insert(path_to, fnv_to, fd_from);
1390
1391         /* Don't free path_to now, since the hash references it */
1392         path_to = NULL;
1393
1394         /* We kept a ref when we removed, and incremented again in insert */
1395         shm_drop(fd_from);
1396         KASSERT(fd_from->shm_refs > 0, ("Expected >0 refs; got: %d\n",
1397             fd_from->shm_refs));
1398
1399         if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) {
1400                 shm_insert(path_from, fnv_from, fd_to);
1401                 path_from = NULL;
1402                 shm_drop(fd_to);
1403                 KASSERT(fd_to->shm_refs > 0, ("Expected >0 refs; got: %d\n",
1404                     fd_to->shm_refs));
1405         }
1406
1407 out_locked:
1408         sx_xunlock(&shm_dict_lock);
1409
1410 out:
1411         free(path_from, M_SHMFD);
1412         free(path_to, M_SHMFD);
1413         return (error);
1414 }
1415
1416 static int
1417 shm_mmap_large(struct shmfd *shmfd, vm_map_t map, vm_offset_t *addr,
1418     vm_size_t size, vm_prot_t prot, vm_prot_t max_prot, int flags,
1419     vm_ooffset_t foff, struct thread *td)
1420 {
1421         struct vmspace *vms;
1422         vm_map_entry_t next_entry, prev_entry;
1423         vm_offset_t align, mask, maxaddr;
1424         int docow, error, rv, try;
1425         bool curmap;
1426
1427         if (shmfd->shm_lp_psind == 0)
1428                 return (EINVAL);
1429
1430         /* MAP_PRIVATE is disabled */
1431         if ((flags & ~(MAP_SHARED | MAP_FIXED | MAP_EXCL |
1432             MAP_NOCORE |
1433 #ifdef MAP_32BIT
1434             MAP_32BIT |
1435 #endif
1436             MAP_ALIGNMENT_MASK)) != 0)
1437                 return (EINVAL);
1438
1439         vms = td->td_proc->p_vmspace;
1440         curmap = map == &vms->vm_map;
1441         if (curmap) {
1442                 error = kern_mmap_racct_check(td, map, size);
1443                 if (error != 0)
1444                         return (error);
1445         }
1446
1447         docow = shmfd->shm_lp_psind << MAP_SPLIT_BOUNDARY_SHIFT;
1448         docow |= MAP_INHERIT_SHARE;
1449         if ((flags & MAP_NOCORE) != 0)
1450                 docow |= MAP_DISABLE_COREDUMP;
1451
1452         mask = pagesizes[shmfd->shm_lp_psind] - 1;
1453         if ((foff & mask) != 0)
1454                 return (EINVAL);
1455         maxaddr = vm_map_max(map);
1456 #ifdef MAP_32BIT
1457         if ((flags & MAP_32BIT) != 0 && maxaddr > MAP_32BIT_MAX_ADDR)
1458                 maxaddr = MAP_32BIT_MAX_ADDR;
1459 #endif
1460         if (size == 0 || (size & mask) != 0 ||
1461             (*addr != 0 && ((*addr & mask) != 0 ||
1462             *addr + size < *addr || *addr + size > maxaddr)))
1463                 return (EINVAL);
1464
1465         align = flags & MAP_ALIGNMENT_MASK;
1466         if (align == 0) {
1467                 align = pagesizes[shmfd->shm_lp_psind];
1468         } else if (align == MAP_ALIGNED_SUPER) {
1469                 if (shmfd->shm_lp_psind != 1)
1470                         return (EINVAL);
1471                 align = pagesizes[1];
1472         } else {
1473                 align >>= MAP_ALIGNMENT_SHIFT;
1474                 align = 1ULL << align;
1475                 /* Also handles overflow. */
1476                 if (align < pagesizes[shmfd->shm_lp_psind])
1477                         return (EINVAL);
1478         }
1479
1480         vm_map_lock(map);
1481         if ((flags & MAP_FIXED) == 0) {
1482                 try = 1;
1483                 if (curmap && (*addr == 0 ||
1484                     (*addr >= round_page((vm_offset_t)vms->vm_taddr) &&
1485                     *addr < round_page((vm_offset_t)vms->vm_daddr +
1486                     lim_max(td, RLIMIT_DATA))))) {
1487                         *addr = roundup2((vm_offset_t)vms->vm_daddr +
1488                             lim_max(td, RLIMIT_DATA),
1489                             pagesizes[shmfd->shm_lp_psind]);
1490                 }
1491 again:
1492                 rv = vm_map_find_aligned(map, addr, size, maxaddr, align);
1493                 if (rv != KERN_SUCCESS) {
1494                         if (try == 1) {
1495                                 try = 2;
1496                                 *addr = vm_map_min(map);
1497                                 if ((*addr & mask) != 0)
1498                                         *addr = (*addr + mask) & mask;
1499                                 goto again;
1500                         }
1501                         goto fail1;
1502                 }
1503         } else if ((flags & MAP_EXCL) == 0) {
1504                 rv = vm_map_delete(map, *addr, *addr + size);
1505                 if (rv != KERN_SUCCESS)
1506                         goto fail1;
1507         } else {
1508                 error = ENOSPC;
1509                 if (vm_map_lookup_entry(map, *addr, &prev_entry))
1510                         goto fail;
1511                 next_entry = vm_map_entry_succ(prev_entry);
1512                 if (next_entry->start < *addr + size)
1513                         goto fail;
1514         }
1515
1516         rv = vm_map_insert(map, shmfd->shm_object, foff, *addr, *addr + size,
1517             prot, max_prot, docow);
1518 fail1:
1519         error = vm_mmap_to_errno(rv);
1520 fail:
1521         vm_map_unlock(map);
1522         return (error);
1523 }
1524
1525 static int
1526 shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize,
1527     vm_prot_t prot, vm_prot_t cap_maxprot, int flags,
1528     vm_ooffset_t foff, struct thread *td)
1529 {
1530         struct shmfd *shmfd;
1531         vm_prot_t maxprot;
1532         int error;
1533         bool writecnt;
1534         void *rl_cookie;
1535
1536         shmfd = fp->f_data;
1537         maxprot = VM_PROT_NONE;
1538
1539         rl_cookie = rangelock_rlock(&shmfd->shm_rl, 0, objsize,
1540             &shmfd->shm_mtx);
1541         /* FREAD should always be set. */
1542         if ((fp->f_flag & FREAD) != 0)
1543                 maxprot |= VM_PROT_EXECUTE | VM_PROT_READ;
1544
1545         /*
1546          * If FWRITE's set, we can allow VM_PROT_WRITE unless it's a shared
1547          * mapping with a write seal applied.  Private mappings are always
1548          * writeable.
1549          */
1550         if ((flags & MAP_SHARED) == 0) {
1551                 cap_maxprot |= VM_PROT_WRITE;
1552                 maxprot |= VM_PROT_WRITE;
1553                 writecnt = false;
1554         } else {
1555                 if ((fp->f_flag & FWRITE) != 0 &&
1556                     (shmfd->shm_seals & F_SEAL_WRITE) == 0)
1557                         maxprot |= VM_PROT_WRITE;
1558
1559                 /*
1560                  * Any mappings from a writable descriptor may be upgraded to
1561                  * VM_PROT_WRITE with mprotect(2), unless a write-seal was
1562                  * applied between the open and subsequent mmap(2).  We want to
1563                  * reject application of a write seal as long as any such
1564                  * mapping exists so that the seal cannot be trivially bypassed.
1565                  */
1566                 writecnt = (maxprot & VM_PROT_WRITE) != 0;
1567                 if (!writecnt && (prot & VM_PROT_WRITE) != 0) {
1568                         error = EACCES;
1569                         goto out;
1570                 }
1571         }
1572         maxprot &= cap_maxprot;
1573
1574         /* See comment in vn_mmap(). */
1575         if (
1576 #ifdef _LP64
1577             objsize > OFF_MAX ||
1578 #endif
1579             foff > OFF_MAX - objsize) {
1580                 error = EINVAL;
1581                 goto out;
1582         }
1583
1584 #ifdef MAC
1585         error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags);
1586         if (error != 0)
1587                 goto out;
1588 #endif
1589
1590         mtx_lock(&shm_timestamp_lock);
1591         vfs_timestamp(&shmfd->shm_atime);
1592         mtx_unlock(&shm_timestamp_lock);
1593         vm_object_reference(shmfd->shm_object);
1594
1595         if (shm_largepage(shmfd)) {
1596                 writecnt = false;
1597                 error = shm_mmap_large(shmfd, map, addr, objsize, prot,
1598                     maxprot, flags, foff, td);
1599         } else {
1600                 if (writecnt) {
1601                         vm_pager_update_writecount(shmfd->shm_object, 0,
1602                             objsize);
1603                 }
1604                 error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags,
1605                     shmfd->shm_object, foff, writecnt, td);
1606         }
1607         if (error != 0) {
1608                 if (writecnt)
1609                         vm_pager_release_writecount(shmfd->shm_object, 0,
1610                             objsize);
1611                 vm_object_deallocate(shmfd->shm_object);
1612         }
1613 out:
1614         rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
1615         return (error);
1616 }
1617
1618 static int
1619 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
1620     struct thread *td)
1621 {
1622         struct shmfd *shmfd;
1623         int error;
1624
1625         error = 0;
1626         shmfd = fp->f_data;
1627         mtx_lock(&shm_timestamp_lock);
1628         /*
1629          * SUSv4 says that x bits of permission need not be affected.
1630          * Be consistent with our shm_open there.
1631          */
1632 #ifdef MAC
1633         error = mac_posixshm_check_setmode(active_cred, shmfd, mode);
1634         if (error != 0)
1635                 goto out;
1636 #endif
1637         error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
1638             VADMIN, active_cred);
1639         if (error != 0)
1640                 goto out;
1641         shmfd->shm_mode = mode & ACCESSPERMS;
1642 out:
1643         mtx_unlock(&shm_timestamp_lock);
1644         return (error);
1645 }
1646
1647 static int
1648 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
1649     struct thread *td)
1650 {
1651         struct shmfd *shmfd;
1652         int error;
1653
1654         error = 0;
1655         shmfd = fp->f_data;
1656         mtx_lock(&shm_timestamp_lock);
1657 #ifdef MAC
1658         error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid);
1659         if (error != 0)
1660                 goto out;
1661 #endif
1662         if (uid == (uid_t)-1)
1663                 uid = shmfd->shm_uid;
1664         if (gid == (gid_t)-1)
1665                  gid = shmfd->shm_gid;
1666         if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) ||
1667             (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) &&
1668             (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN)))
1669                 goto out;
1670         shmfd->shm_uid = uid;
1671         shmfd->shm_gid = gid;
1672 out:
1673         mtx_unlock(&shm_timestamp_lock);
1674         return (error);
1675 }
1676
1677 /*
1678  * Helper routines to allow the backing object of a shared memory file
1679  * descriptor to be mapped in the kernel.
1680  */
1681 int
1682 shm_map(struct file *fp, size_t size, off_t offset, void **memp)
1683 {
1684         struct shmfd *shmfd;
1685         vm_offset_t kva, ofs;
1686         vm_object_t obj;
1687         int rv;
1688
1689         if (fp->f_type != DTYPE_SHM)
1690                 return (EINVAL);
1691         shmfd = fp->f_data;
1692         obj = shmfd->shm_object;
1693         VM_OBJECT_WLOCK(obj);
1694         /*
1695          * XXXRW: This validation is probably insufficient, and subject to
1696          * sign errors.  It should be fixed.
1697          */
1698         if (offset >= shmfd->shm_size ||
1699             offset + size > round_page(shmfd->shm_size)) {
1700                 VM_OBJECT_WUNLOCK(obj);
1701                 return (EINVAL);
1702         }
1703
1704         shmfd->shm_kmappings++;
1705         vm_object_reference_locked(obj);
1706         VM_OBJECT_WUNLOCK(obj);
1707
1708         /* Map the object into the kernel_map and wire it. */
1709         kva = vm_map_min(kernel_map);
1710         ofs = offset & PAGE_MASK;
1711         offset = trunc_page(offset);
1712         size = round_page(size + ofs);
1713         rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0,
1714             VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE,
1715             VM_PROT_READ | VM_PROT_WRITE, 0);
1716         if (rv == KERN_SUCCESS) {
1717                 rv = vm_map_wire(kernel_map, kva, kva + size,
1718                     VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
1719                 if (rv == KERN_SUCCESS) {
1720                         *memp = (void *)(kva + ofs);
1721                         return (0);
1722                 }
1723                 vm_map_remove(kernel_map, kva, kva + size);
1724         } else
1725                 vm_object_deallocate(obj);
1726
1727         /* On failure, drop our mapping reference. */
1728         VM_OBJECT_WLOCK(obj);
1729         shmfd->shm_kmappings--;
1730         VM_OBJECT_WUNLOCK(obj);
1731
1732         return (vm_mmap_to_errno(rv));
1733 }
1734
1735 /*
1736  * We require the caller to unmap the entire entry.  This allows us to
1737  * safely decrement shm_kmappings when a mapping is removed.
1738  */
1739 int
1740 shm_unmap(struct file *fp, void *mem, size_t size)
1741 {
1742         struct shmfd *shmfd;
1743         vm_map_entry_t entry;
1744         vm_offset_t kva, ofs;
1745         vm_object_t obj;
1746         vm_pindex_t pindex;
1747         vm_prot_t prot;
1748         boolean_t wired;
1749         vm_map_t map;
1750         int rv;
1751
1752         if (fp->f_type != DTYPE_SHM)
1753                 return (EINVAL);
1754         shmfd = fp->f_data;
1755         kva = (vm_offset_t)mem;
1756         ofs = kva & PAGE_MASK;
1757         kva = trunc_page(kva);
1758         size = round_page(size + ofs);
1759         map = kernel_map;
1760         rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry,
1761             &obj, &pindex, &prot, &wired);
1762         if (rv != KERN_SUCCESS)
1763                 return (EINVAL);
1764         if (entry->start != kva || entry->end != kva + size) {
1765                 vm_map_lookup_done(map, entry);
1766                 return (EINVAL);
1767         }
1768         vm_map_lookup_done(map, entry);
1769         if (obj != shmfd->shm_object)
1770                 return (EINVAL);
1771         vm_map_remove(map, kva, kva + size);
1772         VM_OBJECT_WLOCK(obj);
1773         KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped"));
1774         shmfd->shm_kmappings--;
1775         VM_OBJECT_WUNLOCK(obj);
1776         return (0);
1777 }
1778
1779 static int
1780 shm_fill_kinfo_locked(struct shmfd *shmfd, struct kinfo_file *kif, bool list)
1781 {
1782         const char *path, *pr_path;
1783         size_t pr_pathlen;
1784         bool visible;
1785
1786         sx_assert(&shm_dict_lock, SA_LOCKED);
1787         kif->kf_type = KF_TYPE_SHM;
1788         kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode;
1789         kif->kf_un.kf_file.kf_file_size = shmfd->shm_size;
1790         if (shmfd->shm_path != NULL) {
1791                 if (shmfd->shm_path != NULL) {
1792                         path = shmfd->shm_path;
1793                         pr_path = curthread->td_ucred->cr_prison->pr_path;
1794                         if (strcmp(pr_path, "/") != 0) {
1795                                 /* Return the jail-rooted pathname. */
1796                                 pr_pathlen = strlen(pr_path);
1797                                 visible = strncmp(path, pr_path, pr_pathlen)
1798                                     == 0 && path[pr_pathlen] == '/';
1799                                 if (list && !visible)
1800                                         return (EPERM);
1801                                 if (visible)
1802                                         path += pr_pathlen;
1803                         }
1804                         strlcpy(kif->kf_path, path, sizeof(kif->kf_path));
1805                 }
1806         }
1807         return (0);
1808 }
1809
1810 static int
1811 shm_fill_kinfo(struct file *fp, struct kinfo_file *kif,
1812     struct filedesc *fdp __unused)
1813 {
1814         int res;
1815
1816         sx_slock(&shm_dict_lock);
1817         res = shm_fill_kinfo_locked(fp->f_data, kif, false);
1818         sx_sunlock(&shm_dict_lock);
1819         return (res);
1820 }
1821
1822 static int
1823 shm_add_seals(struct file *fp, int seals)
1824 {
1825         struct shmfd *shmfd;
1826         void *rl_cookie;
1827         vm_ooffset_t writemappings;
1828         int error, nseals;
1829
1830         error = 0;
1831         shmfd = fp->f_data;
1832         rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
1833             &shmfd->shm_mtx);
1834
1835         /* Even already-set seals should result in EPERM. */
1836         if ((shmfd->shm_seals & F_SEAL_SEAL) != 0) {
1837                 error = EPERM;
1838                 goto out;
1839         }
1840         nseals = seals & ~shmfd->shm_seals;
1841         if ((nseals & F_SEAL_WRITE) != 0) {
1842                 if (shm_largepage(shmfd)) {
1843                         error = ENOTSUP;
1844                         goto out;
1845                 }
1846
1847                 /*
1848                  * The rangelock above prevents writable mappings from being
1849                  * added after we've started applying seals.  The RLOCK here
1850                  * is to avoid torn reads on ILP32 arches as unmapping/reducing
1851                  * writemappings will be done without a rangelock.
1852                  */
1853                 VM_OBJECT_RLOCK(shmfd->shm_object);
1854                 writemappings = shmfd->shm_object->un_pager.swp.writemappings;
1855                 VM_OBJECT_RUNLOCK(shmfd->shm_object);
1856                 /* kmappings are also writable */
1857                 if (writemappings > 0) {
1858                         error = EBUSY;
1859                         goto out;
1860                 }
1861         }
1862         shmfd->shm_seals |= nseals;
1863 out:
1864         rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
1865         return (error);
1866 }
1867
1868 static int
1869 shm_get_seals(struct file *fp, int *seals)
1870 {
1871         struct shmfd *shmfd;
1872
1873         shmfd = fp->f_data;
1874         *seals = shmfd->shm_seals;
1875         return (0);
1876 }
1877
1878 static int
1879 shm_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
1880 {
1881         void *rl_cookie;
1882         struct shmfd *shmfd;
1883         size_t size;
1884         int error;
1885
1886         /* This assumes that the caller already checked for overflow. */
1887         error = 0;
1888         shmfd = fp->f_data;
1889         size = offset + len;
1890
1891         /*
1892          * Just grab the rangelock for the range that we may be attempting to
1893          * grow, rather than blocking read/write for regions we won't be
1894          * touching while this (potential) resize is in progress.  Other
1895          * attempts to resize the shmfd will have to take a write lock from 0 to
1896          * OFF_MAX, so this being potentially beyond the current usable range of
1897          * the shmfd is not necessarily a concern.  If other mechanisms are
1898          * added to grow a shmfd, this may need to be re-evaluated.
1899          */
1900         rl_cookie = rangelock_wlock(&shmfd->shm_rl, offset, size,
1901             &shmfd->shm_mtx);
1902         if (size > shmfd->shm_size)
1903                 error = shm_dotruncate_cookie(shmfd, size, rl_cookie);
1904         rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
1905         /* Translate to posix_fallocate(2) return value as needed. */
1906         if (error == ENOMEM)
1907                 error = ENOSPC;
1908         return (error);
1909 }
1910
1911 static int
1912 sysctl_posix_shm_list(SYSCTL_HANDLER_ARGS)
1913 {
1914         struct shm_mapping *shmm;
1915         struct sbuf sb;
1916         struct kinfo_file kif;
1917         u_long i;
1918         ssize_t curlen;
1919         int error, error2;
1920
1921         sbuf_new_for_sysctl(&sb, NULL, sizeof(struct kinfo_file) * 5, req);
1922         sbuf_clear_flags(&sb, SBUF_INCLUDENUL);
1923         curlen = 0;
1924         error = 0;
1925         sx_slock(&shm_dict_lock);
1926         for (i = 0; i < shm_hash + 1; i++) {
1927                 LIST_FOREACH(shmm, &shm_dictionary[i], sm_link) {
1928                         error = shm_fill_kinfo_locked(shmm->sm_shmfd,
1929                             &kif, true);
1930                         if (error == EPERM) {
1931                                 error = 0;
1932                                 continue;
1933                         }
1934                         if (error != 0)
1935                                 break;
1936                         pack_kinfo(&kif);
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_PRISON | 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 }