sys/kern/uipc_shm.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause
   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 #include "opt_capsicum.h"
  52 #include "opt_ktrace.h"
  53
  54 #include <sys/param.h>
  55 #include <sys/capsicum.h>
  56 #include <sys/conf.h>
  57 #include <sys/fcntl.h>
  58 #include <sys/file.h>
  59 #include <sys/filedesc.h>
  60 #include <sys/filio.h>
  61 #include <sys/fnv_hash.h>
  62 #include <sys/kernel.h>
  63 #include <sys/limits.h>
  64 #include <sys/uio.h>
  65 #include <sys/signal.h>
  66 #include <sys/jail.h>
  67 #include <sys/ktrace.h>
  68 #include <sys/lock.h>
  69 #include <sys/malloc.h>
  70 #include <sys/mman.h>
  71 #include <sys/mutex.h>
  72 #include <sys/priv.h>
  73 #include <sys/proc.h>
  74 #include <sys/refcount.h>
  75 #include <sys/resourcevar.h>
  76 #include <sys/rwlock.h>
  77 #include <sys/sbuf.h>
  78 #include <sys/stat.h>
  79 #include <sys/syscallsubr.h>
  80 #include <sys/sysctl.h>
  81 #include <sys/sysproto.h>
  82 #include <sys/systm.h>
  83 #include <sys/sx.h>
  84 #include <sys/time.h>
  85 #include <sys/vmmeter.h>
  86 #include <sys/vnode.h>
  87 #include <sys/unistd.h>
  88 #include <sys/user.h>
  89
  90 #include <security/audit/audit.h>
  91 #include <security/mac/mac_framework.h>
  92
  93 #include <vm/vm.h>
  94 #include <vm/vm_param.h>
  95 #include <vm/pmap.h>
  96 #include <vm/vm_extern.h>
  97 #include <vm/vm_map.h>
  98 #include <vm/vm_kern.h>
  99 #include <vm/vm_object.h>
 100 #include <vm/vm_page.h>
 101 #include <vm/vm_pageout.h>
 102 #include <vm/vm_pager.h>
 103 #include <vm/swap_pager.h>
 104
 105 struct shm_mapping {
 106         char            *sm_path;
 107         Fnv32_t         sm_fnv;
 108         struct shmfd    *sm_shmfd;
 109         LIST_ENTRY(shm_mapping) sm_link;
 110 };
 111
 112 static MALLOC_DEFINE(M_SHMFD, "shmfd", "shared memory file descriptor");
 113 static LIST_HEAD(, shm_mapping) *shm_dictionary;
 114 static struct sx shm_dict_lock;
 115 static struct mtx shm_timestamp_lock;
 116 static u_long shm_hash;
 117 static struct unrhdr64 shm_ino_unr;
 118 static dev_t shm_dev_ino;
 119
 120 #define SHM_HASH(fnv)   (&shm_dictionary[(fnv) & shm_hash])
 121
 122 static void     shm_init(void *arg);
 123 static void     shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd);
 124 static struct shmfd *shm_lookup(char *path, Fnv32_t fnv);
 125 static int      shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred);
 126 static void     shm_doremove(struct shm_mapping *map);
 127 static int      shm_dotruncate_cookie(struct shmfd *shmfd, off_t length,
 128     void *rl_cookie);
 129 static int      shm_dotruncate_locked(struct shmfd *shmfd, off_t length,
 130     void *rl_cookie);
 131 static int      shm_copyin_path(struct thread *td, const char *userpath_in,
 132     char **path_out);
 133 static int      shm_deallocate(struct shmfd *shmfd, off_t *offset,
 134     off_t *length, int flags);
 135
 136 static fo_rdwr_t        shm_read;
 137 static fo_rdwr_t        shm_write;
 138 static fo_truncate_t    shm_truncate;
 139 static fo_ioctl_t       shm_ioctl;
 140 static fo_stat_t        shm_stat;
 141 static fo_close_t       shm_close;
 142 static fo_chmod_t       shm_chmod;
 143 static fo_chown_t       shm_chown;
 144 static fo_seek_t        shm_seek;
 145 static fo_fill_kinfo_t  shm_fill_kinfo;
 146 static fo_mmap_t        shm_mmap;
 147 static fo_get_seals_t   shm_get_seals;
 148 static fo_add_seals_t   shm_add_seals;
 149 static fo_fallocate_t   shm_fallocate;
 150 static fo_fspacectl_t   shm_fspacectl;
 151
 152 /* File descriptor operations. */
 153 struct fileops shm_ops = {
 154         .fo_read = shm_read,
 155         .fo_write = shm_write,
 156         .fo_truncate = shm_truncate,
 157         .fo_ioctl = shm_ioctl,
 158         .fo_poll = invfo_poll,
 159         .fo_kqfilter = invfo_kqfilter,
 160         .fo_stat = shm_stat,
 161         .fo_close = shm_close,
 162         .fo_chmod = shm_chmod,
 163         .fo_chown = shm_chown,
 164         .fo_sendfile = vn_sendfile,
 165         .fo_seek = shm_seek,
 166         .fo_fill_kinfo = shm_fill_kinfo,
 167         .fo_mmap = shm_mmap,
 168         .fo_get_seals = shm_get_seals,
 169         .fo_add_seals = shm_add_seals,
 170         .fo_fallocate = shm_fallocate,
 171         .fo_fspacectl = shm_fspacectl,
 172         .fo_cmp = file_kcmp_generic,
 173         .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE,
 174 };
 175
 176 FEATURE(posix_shm, "POSIX shared memory");
 177
 178 static SYSCTL_NODE(_vm, OID_AUTO, largepages, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
 179     "");
 180
 181 static int largepage_reclaim_tries = 1;
 182 SYSCTL_INT(_vm_largepages, OID_AUTO, reclaim_tries,
 183     CTLFLAG_RWTUN, &largepage_reclaim_tries, 0,
 184     "Number of contig reclaims before giving up for default alloc policy");
 185
 186 #define shm_rangelock_unlock(shmfd, cookie)                             \
 187         rangelock_unlock(&(shmfd)->shm_rl, (cookie), &(shmfd)->shm_mtx)
 188 #define shm_rangelock_rlock(shmfd, start, end)                          \
 189         rangelock_rlock(&(shmfd)->shm_rl, (start), (end), &(shmfd)->shm_mtx)
 190 #define shm_rangelock_tryrlock(shmfd, start, end)                       \
 191         rangelock_tryrlock(&(shmfd)->shm_rl, (start), (end), &(shmfd)->shm_mtx)
 192 #define shm_rangelock_wlock(shmfd, start, end)                          \
 193         rangelock_wlock(&(shmfd)->shm_rl, (start), (end), &(shmfd)->shm_mtx)
 194
 195 static int
 196 uiomove_object_page(vm_object_t obj, size_t len, struct uio *uio)
 197 {
 198         vm_page_t m;
 199         vm_pindex_t idx;
 200         size_t tlen;
 201         int error, offset, rv;
 202
 203         idx = OFF_TO_IDX(uio->uio_offset);
 204         offset = uio->uio_offset & PAGE_MASK;
 205         tlen = MIN(PAGE_SIZE - offset, len);
 206
 207         rv = vm_page_grab_valid_unlocked(&m, obj, idx,
 208             VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY | VM_ALLOC_NOCREAT);
 209         if (rv == VM_PAGER_OK)
 210                 goto found;
 211
 212         /*
 213          * Read I/O without either a corresponding resident page or swap
 214          * page: use zero_region.  This is intended to avoid instantiating
 215          * pages on read from a sparse region.
 216          */
 217         VM_OBJECT_WLOCK(obj);
 218         m = vm_page_lookup(obj, idx);
 219         if (uio->uio_rw == UIO_READ && m == NULL &&
 220             !vm_pager_has_page(obj, idx, NULL, NULL)) {
 221                 VM_OBJECT_WUNLOCK(obj);
 222                 return (uiomove(__DECONST(void *, zero_region), tlen, uio));
 223         }
 224
 225         /*
 226          * Although the tmpfs vnode lock is held here, it is
 227          * nonetheless safe to sleep waiting for a free page.  The
 228          * pageout daemon does not need to acquire the tmpfs vnode
 229          * lock to page out tobj's pages because tobj is a OBJT_SWAP
 230          * type object.
 231          */
 232         rv = vm_page_grab_valid(&m, obj, idx,
 233             VM_ALLOC_NORMAL | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY);
 234         if (rv != VM_PAGER_OK) {
 235                 VM_OBJECT_WUNLOCK(obj);
 236                 if (bootverbose) {
 237                         printf("uiomove_object: vm_obj %p idx %jd "
 238                             "pager error %d\n", obj, idx, rv);
 239                 }
 240                 return (rv == VM_PAGER_AGAIN ? ENOSPC : EIO);
 241         }
 242         VM_OBJECT_WUNLOCK(obj);
 243
 244 found:
 245         error = uiomove_fromphys(&m, offset, tlen, uio);
 246         if (uio->uio_rw == UIO_WRITE && error == 0)
 247                 vm_page_set_dirty(m);
 248         vm_page_activate(m);
 249         vm_page_sunbusy(m);
 250
 251         return (error);
 252 }
 253
 254 int
 255 uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio)
 256 {
 257         ssize_t resid;
 258         size_t len;
 259         int error;
 260
 261         error = 0;
 262         while ((resid = uio->uio_resid) > 0) {
 263                 if (obj_size <= uio->uio_offset)
 264                         break;
 265                 len = MIN(obj_size - uio->uio_offset, resid);
 266                 if (len == 0)
 267                         break;
 268                 error = uiomove_object_page(obj, len, uio);
 269                 if (error != 0 || resid == uio->uio_resid)
 270                         break;
 271         }
 272         return (error);
 273 }
 274
 275 static u_long count_largepages[MAXPAGESIZES];
 276
 277 static int
 278 shm_largepage_phys_populate(vm_object_t object, vm_pindex_t pidx,
 279     int fault_type, vm_prot_t max_prot, vm_pindex_t *first, vm_pindex_t *last)
 280 {
 281         vm_page_t m __diagused;
 282         int psind;
 283
 284         psind = object->un_pager.phys.data_val;
 285         if (psind == 0 || pidx >= object->size)
 286                 return (VM_PAGER_FAIL);
 287         *first = rounddown2(pidx, pagesizes[psind] / PAGE_SIZE);
 288
 289         /*
 290          * We only busy the first page in the superpage run.  It is
 291          * useless to busy whole run since we only remove full
 292          * superpage, and it takes too long to busy e.g. 512 * 512 ==
 293          * 262144 pages constituing 1G amd64 superage.
 294          */
 295         m = vm_page_grab(object, *first, VM_ALLOC_NORMAL | VM_ALLOC_NOCREAT);
 296         MPASS(m != NULL);
 297
 298         *last = *first + atop(pagesizes[psind]) - 1;
 299         return (VM_PAGER_OK);
 300 }
 301
 302 static boolean_t
 303 shm_largepage_phys_haspage(vm_object_t object, vm_pindex_t pindex,
 304     int *before, int *after)
 305 {
 306         int psind;
 307
 308         psind = object->un_pager.phys.data_val;
 309         if (psind == 0 || pindex >= object->size)
 310                 return (FALSE);
 311         if (before != NULL) {
 312                 *before = pindex - rounddown2(pindex, pagesizes[psind] /
 313                     PAGE_SIZE);
 314         }
 315         if (after != NULL) {
 316                 *after = roundup2(pindex, pagesizes[psind] / PAGE_SIZE) -
 317                     pindex;
 318         }
 319         return (TRUE);
 320 }
 321
 322 static void
 323 shm_largepage_phys_ctor(vm_object_t object, vm_prot_t prot,
 324     vm_ooffset_t foff, struct ucred *cred)
 325 {
 326 }
 327
 328 static void
 329 shm_largepage_phys_dtor(vm_object_t object)
 330 {
 331         int psind;
 332
 333         psind = object->un_pager.phys.data_val;
 334         if (psind != 0) {
 335                 atomic_subtract_long(&count_largepages[psind],
 336                     object->size / (pagesizes[psind] / PAGE_SIZE));
 337                 vm_wire_sub(object->size);
 338         } else {
 339                 KASSERT(object->size == 0,
 340                     ("largepage phys obj %p not initialized bit size %#jx > 0",
 341                     object, (uintmax_t)object->size));
 342         }
 343 }
 344
 345 static const struct phys_pager_ops shm_largepage_phys_ops = {
 346         .phys_pg_populate =     shm_largepage_phys_populate,
 347         .phys_pg_haspage =      shm_largepage_phys_haspage,
 348         .phys_pg_ctor =         shm_largepage_phys_ctor,
 349         .phys_pg_dtor =         shm_largepage_phys_dtor,
 350 };
 351
 352 bool
 353 shm_largepage(struct shmfd *shmfd)
 354 {
 355         return (shmfd->shm_object->type == OBJT_PHYS);
 356 }
 357
 358 static void
 359 shm_pager_freespace(vm_object_t obj, vm_pindex_t start, vm_size_t size)
 360 {
 361         struct shmfd *shm;
 362         vm_size_t c;
 363
 364         swap_pager_freespace(obj, start, size, &c);
 365         if (c == 0)
 366                 return;
 367
 368         shm = obj->un_pager.swp.swp_priv;
 369         if (shm == NULL)
 370                 return;
 371         KASSERT(shm->shm_pages >= c,
 372             ("shm %p pages %jd free %jd", shm,
 373             (uintmax_t)shm->shm_pages, (uintmax_t)c));
 374         shm->shm_pages -= c;
 375 }
 376
 377 static void
 378 shm_page_inserted(vm_object_t obj, vm_page_t m)
 379 {
 380         struct shmfd *shm;
 381
 382         shm = obj->un_pager.swp.swp_priv;
 383         if (shm == NULL)
 384                 return;
 385         if (!vm_pager_has_page(obj, m->pindex, NULL, NULL))
 386                 shm->shm_pages += 1;
 387 }
 388
 389 static void
 390 shm_page_removed(vm_object_t obj, vm_page_t m)
 391 {
 392         struct shmfd *shm;
 393
 394         shm = obj->un_pager.swp.swp_priv;
 395         if (shm == NULL)
 396                 return;
 397         if (!vm_pager_has_page(obj, m->pindex, NULL, NULL)) {
 398                 KASSERT(shm->shm_pages >= 1,
 399                     ("shm %p pages %jd free 1", shm,
 400                     (uintmax_t)shm->shm_pages));
 401                 shm->shm_pages -= 1;
 402         }
 403 }
 404
 405 static struct pagerops shm_swap_pager_ops = {
 406         .pgo_kvme_type = KVME_TYPE_SWAP,
 407         .pgo_freespace = shm_pager_freespace,
 408         .pgo_page_inserted = shm_page_inserted,
 409         .pgo_page_removed = shm_page_removed,
 410 };
 411 static int shmfd_pager_type = -1;
 412
 413 static int
 414 shm_seek(struct file *fp, off_t offset, int whence, struct thread *td)
 415 {
 416         struct shmfd *shmfd;
 417         off_t foffset;
 418         int error;
 419
 420         shmfd = fp->f_data;
 421         foffset = foffset_lock(fp, 0);
 422         error = 0;
 423         switch (whence) {
 424         case L_INCR:
 425                 if (foffset < 0 ||
 426                     (offset > 0 && foffset > OFF_MAX - offset)) {
 427                         error = EOVERFLOW;
 428                         break;
 429                 }
 430                 offset += foffset;
 431                 break;
 432         case L_XTND:
 433                 if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) {
 434                         error = EOVERFLOW;
 435                         break;
 436                 }
 437                 offset += shmfd->shm_size;
 438                 break;
 439         case L_SET:
 440                 break;
 441         default:
 442                 error = EINVAL;
 443         }
 444         if (error == 0) {
 445                 if (offset < 0 || offset > shmfd->shm_size)
 446                         error = EINVAL;
 447                 else
 448                         td->td_uretoff.tdu_off = offset;
 449         }
 450         foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
 451         return (error);
 452 }
 453
 454 static int
 455 shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
 456     int flags, struct thread *td)
 457 {
 458         struct shmfd *shmfd;
 459         void *rl_cookie;
 460         int error;
 461
 462         shmfd = fp->f_data;
 463 #ifdef MAC
 464         error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd);
 465         if (error)
 466                 return (error);
 467 #endif
 468         foffset_lock_uio(fp, uio, flags);
 469         rl_cookie = shm_rangelock_rlock(shmfd, uio->uio_offset,
 470             uio->uio_offset + uio->uio_resid);
 471         error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
 472         shm_rangelock_unlock(shmfd, rl_cookie);
 473         foffset_unlock_uio(fp, uio, flags);
 474         return (error);
 475 }
 476
 477 static int
 478 shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
 479     int flags, struct thread *td)
 480 {
 481         struct shmfd *shmfd;
 482         void *rl_cookie;
 483         int error;
 484         off_t size;
 485
 486         shmfd = fp->f_data;
 487 #ifdef MAC
 488         error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd);
 489         if (error)
 490                 return (error);
 491 #endif
 492         if (shm_largepage(shmfd) && shmfd->shm_lp_psind == 0)
 493                 return (EINVAL);
 494         foffset_lock_uio(fp, uio, flags);
 495         if (uio->uio_resid > OFF_MAX - uio->uio_offset) {
 496                 /*
 497                  * Overflow is only an error if we're supposed to expand on
 498                  * write.  Otherwise, we'll just truncate the write to the
 499                  * size of the file, which can only grow up to OFF_MAX.
 500                  */
 501                 if ((shmfd->shm_flags & SHM_GROW_ON_WRITE) != 0) {
 502                         foffset_unlock_uio(fp, uio, flags);
 503                         return (EFBIG);
 504                 }
 505
 506                 size = shmfd->shm_size;
 507         } else {
 508                 size = uio->uio_offset + uio->uio_resid;
 509         }
 510         if ((flags & FOF_OFFSET) == 0)
 511                 rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX);
 512         else
 513                 rl_cookie = shm_rangelock_wlock(shmfd, uio->uio_offset, size);
 514         if ((shmfd->shm_seals & F_SEAL_WRITE) != 0) {
 515                 error = EPERM;
 516         } else {
 517                 error = 0;
 518                 if ((shmfd->shm_flags & SHM_GROW_ON_WRITE) != 0 &&
 519                     size > shmfd->shm_size) {
 520                         error = shm_dotruncate_cookie(shmfd, size, rl_cookie);
 521                 }
 522                 if (error == 0)
 523                         error = uiomove_object(shmfd->shm_object,
 524                             shmfd->shm_size, uio);
 525         }
 526         shm_rangelock_unlock(shmfd, rl_cookie);
 527         foffset_unlock_uio(fp, uio, flags);
 528         return (error);
 529 }
 530
 531 static int
 532 shm_truncate(struct file *fp, off_t length, struct ucred *active_cred,
 533     struct thread *td)
 534 {
 535         struct shmfd *shmfd;
 536 #ifdef MAC
 537         int error;
 538 #endif
 539
 540         shmfd = fp->f_data;
 541 #ifdef MAC
 542         error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd);
 543         if (error)
 544                 return (error);
 545 #endif
 546         return (shm_dotruncate(shmfd, length));
 547 }
 548
 549 int
 550 shm_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
 551     struct thread *td)
 552 {
 553         struct shmfd *shmfd;
 554         struct shm_largepage_conf *conf;
 555         void *rl_cookie;
 556
 557         shmfd = fp->f_data;
 558         switch (com) {
 559         case FIONBIO:
 560         case FIOASYNC:
 561                 /*
 562                  * Allow fcntl(fd, F_SETFL, O_NONBLOCK) to work,
 563                  * just like it would on an unlinked regular file
 564                  */
 565                 return (0);
 566         case FIOSSHMLPGCNF:
 567                 if (!shm_largepage(shmfd))
 568                         return (ENOTTY);
 569                 conf = data;
 570                 if (shmfd->shm_lp_psind != 0 &&
 571                     conf->psind != shmfd->shm_lp_psind)
 572                         return (EINVAL);
 573                 if (conf->psind <= 0 || conf->psind >= MAXPAGESIZES ||
 574                     pagesizes[conf->psind] == 0)
 575                         return (EINVAL);
 576                 if (conf->alloc_policy != SHM_LARGEPAGE_ALLOC_DEFAULT &&
 577                     conf->alloc_policy != SHM_LARGEPAGE_ALLOC_NOWAIT &&
 578                     conf->alloc_policy != SHM_LARGEPAGE_ALLOC_HARD)
 579                         return (EINVAL);
 580
 581                 rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX);
 582                 shmfd->shm_lp_psind = conf->psind;
 583                 shmfd->shm_lp_alloc_policy = conf->alloc_policy;
 584                 shmfd->shm_object->un_pager.phys.data_val = conf->psind;
 585                 shm_rangelock_unlock(shmfd, rl_cookie);
 586                 return (0);
 587         case FIOGSHMLPGCNF:
 588                 if (!shm_largepage(shmfd))
 589                         return (ENOTTY);
 590                 conf = data;
 591                 rl_cookie = shm_rangelock_rlock(shmfd, 0, OFF_MAX);
 592                 conf->psind = shmfd->shm_lp_psind;
 593                 conf->alloc_policy = shmfd->shm_lp_alloc_policy;
 594                 shm_rangelock_unlock(shmfd, rl_cookie);
 595                 return (0);
 596         default:
 597                 return (ENOTTY);
 598         }
 599 }
 600
 601 static int
 602 shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred)
 603 {
 604         struct shmfd *shmfd;
 605 #ifdef MAC
 606         int error;
 607 #endif
 608
 609         shmfd = fp->f_data;
 610
 611 #ifdef MAC
 612         error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd);
 613         if (error)
 614                 return (error);
 615 #endif
 616
 617         /*
 618          * Attempt to return sanish values for fstat() on a memory file
 619          * descriptor.
 620          */
 621         bzero(sb, sizeof(*sb));
 622         sb->st_blksize = PAGE_SIZE;
 623         sb->st_size = shmfd->shm_size;
 624         mtx_lock(&shm_timestamp_lock);
 625         sb->st_atim = shmfd->shm_atime;
 626         sb->st_ctim = shmfd->shm_ctime;
 627         sb->st_mtim = shmfd->shm_mtime;
 628         sb->st_birthtim = shmfd->shm_birthtime;
 629         sb->st_mode = S_IFREG | shmfd->shm_mode;                /* XXX */
 630         sb->st_uid = shmfd->shm_uid;
 631         sb->st_gid = shmfd->shm_gid;
 632         mtx_unlock(&shm_timestamp_lock);
 633         sb->st_dev = shm_dev_ino;
 634         sb->st_ino = shmfd->shm_ino;
 635         sb->st_nlink = shmfd->shm_object->ref_count;
 636         if (shm_largepage(shmfd)) {
 637                 sb->st_blocks = shmfd->shm_object->size /
 638                     (pagesizes[shmfd->shm_lp_psind] >> PAGE_SHIFT);
 639         } else {
 640                 sb->st_blocks = shmfd->shm_pages;
 641         }
 642
 643         return (0);
 644 }
 645
 646 static int
 647 shm_close(struct file *fp, struct thread *td)
 648 {
 649         struct shmfd *shmfd;
 650
 651         shmfd = fp->f_data;
 652         fp->f_data = NULL;
 653         shm_drop(shmfd);
 654
 655         return (0);
 656 }
 657
 658 static int
 659 shm_copyin_path(struct thread *td, const char *userpath_in, char **path_out) {
 660         int error;
 661         char *path;
 662         const char *pr_path;
 663         size_t pr_pathlen;
 664
 665         path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK);
 666         pr_path = td->td_ucred->cr_prison->pr_path;
 667
 668         /* Construct a full pathname for jailed callers. */
 669         pr_pathlen = strcmp(pr_path, "/") ==
 670             0 ? 0 : strlcpy(path, pr_path, MAXPATHLEN);
 671         error = copyinstr(userpath_in, path + pr_pathlen,
 672             MAXPATHLEN - pr_pathlen, NULL);
 673         if (error != 0)
 674                 goto out;
 675
 676 #ifdef KTRACE
 677         if (KTRPOINT(curthread, KTR_NAMEI))
 678                 ktrnamei(path);
 679 #endif
 680
 681         /* Require paths to start with a '/' character. */
 682         if (path[pr_pathlen] != '/') {
 683                 error = EINVAL;
 684                 goto out;
 685         }
 686
 687         *path_out = path;
 688
 689 out:
 690         if (error != 0)
 691                 free(path, M_SHMFD);
 692
 693         return (error);
 694 }
 695
 696 static int
 697 shm_partial_page_invalidate(vm_object_t object, vm_pindex_t idx, int base,
 698     int end)
 699 {
 700         vm_page_t m;
 701         int rv;
 702
 703         VM_OBJECT_ASSERT_WLOCKED(object);
 704         KASSERT(base >= 0, ("%s: base %d", __func__, base));
 705         KASSERT(end - base <= PAGE_SIZE, ("%s: base %d end %d", __func__, base,
 706             end));
 707
 708 retry:
 709         m = vm_page_grab(object, idx, VM_ALLOC_NOCREAT);
 710         if (m != NULL) {
 711                 MPASS(vm_page_all_valid(m));
 712         } else if (vm_pager_has_page(object, idx, NULL, NULL)) {
 713                 m = vm_page_alloc(object, idx,
 714                     VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL);
 715                 if (m == NULL)
 716                         goto retry;
 717                 vm_object_pip_add(object, 1);
 718                 VM_OBJECT_WUNLOCK(object);
 719                 rv = vm_pager_get_pages(object, &m, 1, NULL, NULL);
 720                 VM_OBJECT_WLOCK(object);
 721                 vm_object_pip_wakeup(object);
 722                 if (rv == VM_PAGER_OK) {
 723                         /*
 724                          * Since the page was not resident, and therefore not
 725                          * recently accessed, immediately enqueue it for
 726                          * asynchronous laundering.  The current operation is
 727                          * not regarded as an access.
 728                          */
 729                         vm_page_launder(m);
 730                 } else {
 731                         vm_page_free(m);
 732                         VM_OBJECT_WUNLOCK(object);
 733                         return (EIO);
 734                 }
 735         }
 736         if (m != NULL) {
 737                 pmap_zero_page_area(m, base, end - base);
 738                 KASSERT(vm_page_all_valid(m), ("%s: page %p is invalid",
 739                     __func__, m));
 740                 vm_page_set_dirty(m);
 741                 vm_page_xunbusy(m);
 742         }
 743
 744         return (0);
 745 }
 746
 747 static int
 748 shm_dotruncate_locked(struct shmfd *shmfd, off_t length, void *rl_cookie)
 749 {
 750         vm_object_t object;
 751         vm_pindex_t nobjsize;
 752         vm_ooffset_t delta;
 753         int base, error;
 754
 755         KASSERT(length >= 0, ("shm_dotruncate: length < 0"));
 756         object = shmfd->shm_object;
 757         VM_OBJECT_ASSERT_WLOCKED(object);
 758         rangelock_cookie_assert(rl_cookie, RA_WLOCKED);
 759         if (length == shmfd->shm_size)
 760                 return (0);
 761         nobjsize = OFF_TO_IDX(length + PAGE_MASK);
 762
 763         /* Are we shrinking?  If so, trim the end. */
 764         if (length < shmfd->shm_size) {
 765                 if ((shmfd->shm_seals & F_SEAL_SHRINK) != 0)
 766                         return (EPERM);
 767
 768                 /*
 769                  * Disallow any requests to shrink the size if this
 770                  * object is mapped into the kernel.
 771                  */
 772                 if (shmfd->shm_kmappings > 0)
 773                         return (EBUSY);
 774
 775                 /*
 776                  * Zero the truncated part of the last page.
 777                  */
 778                 base = length & PAGE_MASK;
 779                 if (base != 0) {
 780                         error = shm_partial_page_invalidate(object,
 781                             OFF_TO_IDX(length), base, PAGE_SIZE);
 782                         if (error)
 783                                 return (error);
 784                 }
 785                 delta = IDX_TO_OFF(object->size - nobjsize);
 786
 787                 if (nobjsize < object->size)
 788                         vm_object_page_remove(object, nobjsize, object->size,
 789                             0);
 790
 791                 /* Free the swap accounted for shm */
 792                 swap_release_by_cred(delta, object->cred);
 793                 object->charge -= delta;
 794         } else {
 795                 if ((shmfd->shm_seals & F_SEAL_GROW) != 0)
 796                         return (EPERM);
 797
 798                 /* Try to reserve additional swap space. */
 799                 delta = IDX_TO_OFF(nobjsize - object->size);
 800                 if (!swap_reserve_by_cred(delta, object->cred))
 801                         return (ENOMEM);
 802                 object->charge += delta;
 803         }
 804         shmfd->shm_size = length;
 805         mtx_lock(&shm_timestamp_lock);
 806         vfs_timestamp(&shmfd->shm_ctime);
 807         shmfd->shm_mtime = shmfd->shm_ctime;
 808         mtx_unlock(&shm_timestamp_lock);
 809         object->size = nobjsize;
 810         return (0);
 811 }
 812
 813 static int
 814 shm_dotruncate_largepage(struct shmfd *shmfd, off_t length, void *rl_cookie)
 815 {
 816         vm_object_t object;
 817         vm_page_t m;
 818         vm_pindex_t newobjsz;
 819         vm_pindex_t oldobjsz __unused;
 820         int aflags, error, i, psind, try;
 821
 822         KASSERT(length >= 0, ("shm_dotruncate: length < 0"));
 823         object = shmfd->shm_object;
 824         VM_OBJECT_ASSERT_WLOCKED(object);
 825         rangelock_cookie_assert(rl_cookie, RA_WLOCKED);
 826
 827         oldobjsz = object->size;
 828         newobjsz = OFF_TO_IDX(length);
 829         if (length == shmfd->shm_size)
 830                 return (0);
 831         psind = shmfd->shm_lp_psind;
 832         if (psind == 0 && length != 0)
 833                 return (EINVAL);
 834         if ((length & (pagesizes[psind] - 1)) != 0)
 835                 return (EINVAL);
 836
 837         if (length < shmfd->shm_size) {
 838                 if ((shmfd->shm_seals & F_SEAL_SHRINK) != 0)
 839                         return (EPERM);
 840                 if (shmfd->shm_kmappings > 0)
 841                         return (EBUSY);
 842                 return (ENOTSUP);       /* Pages are unmanaged. */
 843 #if 0
 844                 vm_object_page_remove(object, newobjsz, oldobjsz, 0);
 845                 object->size = newobjsz;
 846                 shmfd->shm_size = length;
 847                 return (0);
 848 #endif
 849         }
 850
 851         if ((shmfd->shm_seals & F_SEAL_GROW) != 0)
 852                 return (EPERM);
 853
 854         aflags = VM_ALLOC_NORMAL | VM_ALLOC_ZERO;
 855         if (shmfd->shm_lp_alloc_policy == SHM_LARGEPAGE_ALLOC_NOWAIT)
 856                 aflags |= VM_ALLOC_WAITFAIL;
 857         try = 0;
 858
 859         /*
 860          * Extend shmfd and object, keeping all already fully
 861          * allocated large pages intact even on error, because dropped
 862          * object lock might allowed mapping of them.
 863          */
 864         while (object->size < newobjsz) {
 865                 m = vm_page_alloc_contig(object, object->size, aflags,
 866                     pagesizes[psind] / PAGE_SIZE, 0, ~0,
 867                     pagesizes[psind], 0,
 868                     VM_MEMATTR_DEFAULT);
 869                 if (m == NULL) {
 870                         VM_OBJECT_WUNLOCK(object);
 871                         if (shmfd->shm_lp_alloc_policy ==
 872                             SHM_LARGEPAGE_ALLOC_NOWAIT ||
 873                             (shmfd->shm_lp_alloc_policy ==
 874                             SHM_LARGEPAGE_ALLOC_DEFAULT &&
 875                             try >= largepage_reclaim_tries)) {
 876                                 VM_OBJECT_WLOCK(object);
 877                                 return (ENOMEM);
 878                         }
 879                         error = vm_page_reclaim_contig(aflags,
 880                             pagesizes[psind] / PAGE_SIZE, 0, ~0,
 881                             pagesizes[psind], 0);
 882                         if (error == ENOMEM)
 883                                 error = vm_wait_intr(object);
 884                         if (error != 0) {
 885                                 VM_OBJECT_WLOCK(object);
 886                                 return (error);
 887                         }
 888                         try++;
 889                         VM_OBJECT_WLOCK(object);
 890                         continue;
 891                 }
 892                 try = 0;
 893                 for (i = 0; i < pagesizes[psind] / PAGE_SIZE; i++) {
 894                         if ((m[i].flags & PG_ZERO) == 0)
 895                                 pmap_zero_page(&m[i]);
 896                         vm_page_valid(&m[i]);
 897                         vm_page_xunbusy(&m[i]);
 898                 }
 899                 object->size += OFF_TO_IDX(pagesizes[psind]);
 900                 shmfd->shm_size += pagesizes[psind];
 901                 atomic_add_long(&count_largepages[psind], 1);
 902                 vm_wire_add(atop(pagesizes[psind]));
 903         }
 904         return (0);
 905 }
 906
 907 static int
 908 shm_dotruncate_cookie(struct shmfd *shmfd, off_t length, void *rl_cookie)
 909 {
 910         int error;
 911
 912         VM_OBJECT_WLOCK(shmfd->shm_object);
 913         error = shm_largepage(shmfd) ? shm_dotruncate_largepage(shmfd,
 914             length, rl_cookie) : shm_dotruncate_locked(shmfd, length,
 915             rl_cookie);
 916         VM_OBJECT_WUNLOCK(shmfd->shm_object);
 917         return (error);
 918 }
 919
 920 int
 921 shm_dotruncate(struct shmfd *shmfd, off_t length)
 922 {
 923         void *rl_cookie;
 924         int error;
 925
 926         rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX);
 927         error = shm_dotruncate_cookie(shmfd, length, rl_cookie);
 928         shm_rangelock_unlock(shmfd, rl_cookie);
 929         return (error);
 930 }
 931
 932 /*
 933  * shmfd object management including creation and reference counting
 934  * routines.
 935  */
 936 struct shmfd *
 937 shm_alloc(struct ucred *ucred, mode_t mode, bool largepage)
 938 {
 939         struct shmfd *shmfd;
 940         vm_object_t obj;
 941
 942         shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO);
 943         shmfd->shm_size = 0;
 944         shmfd->shm_uid = ucred->cr_uid;
 945         shmfd->shm_gid = ucred->cr_gid;
 946         shmfd->shm_mode = mode;
 947         if (largepage) {
 948                 shmfd->shm_object = phys_pager_allocate(NULL,
 949                     &shm_largepage_phys_ops, NULL, shmfd->shm_size,
 950                     VM_PROT_DEFAULT, 0, ucred);
 951                 shmfd->shm_lp_alloc_policy = SHM_LARGEPAGE_ALLOC_DEFAULT;
 952         } else {
 953                 obj = vm_pager_allocate(shmfd_pager_type, NULL,
 954                     shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred);
 955                 VM_OBJECT_WLOCK(obj);
 956                 obj->un_pager.swp.swp_priv = shmfd;
 957                 VM_OBJECT_WUNLOCK(obj);
 958                 shmfd->shm_object = obj;
 959         }
 960         KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate"));
 961         vfs_timestamp(&shmfd->shm_birthtime);
 962         shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime =
 963             shmfd->shm_birthtime;
 964         shmfd->shm_ino = alloc_unr64(&shm_ino_unr);
 965         refcount_init(&shmfd->shm_refs, 1);
 966         mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF);
 967         rangelock_init(&shmfd->shm_rl);
 968 #ifdef MAC
 969         mac_posixshm_init(shmfd);
 970         mac_posixshm_create(ucred, shmfd);
 971 #endif
 972
 973         return (shmfd);
 974 }
 975
 976 struct shmfd *
 977 shm_hold(struct shmfd *shmfd)
 978 {
 979
 980         refcount_acquire(&shmfd->shm_refs);
 981         return (shmfd);
 982 }
 983
 984 void
 985 shm_drop(struct shmfd *shmfd)
 986 {
 987         vm_object_t obj;
 988
 989         if (refcount_release(&shmfd->shm_refs)) {
 990 #ifdef MAC
 991                 mac_posixshm_destroy(shmfd);
 992 #endif
 993                 rangelock_destroy(&shmfd->shm_rl);
 994                 mtx_destroy(&shmfd->shm_mtx);
 995                 obj = shmfd->shm_object;
 996                 if (!shm_largepage(shmfd)) {
 997                         VM_OBJECT_WLOCK(obj);
 998                         obj->un_pager.swp.swp_priv = NULL;
 999                         VM_OBJECT_WUNLOCK(obj);
1000                 }
1001                 vm_object_deallocate(obj);
1002                 free(shmfd, M_SHMFD);
1003         }
1004 }
1005
1006 /*
1007  * Determine if the credentials have sufficient permissions for a
1008  * specified combination of FREAD and FWRITE.
1009  */
1010 int
1011 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags)
1012 {
1013         accmode_t accmode;
1014         int error;
1015
1016         accmode = 0;
1017         if (flags & FREAD)
1018                 accmode |= VREAD;
1019         if (flags & FWRITE)
1020                 accmode |= VWRITE;
1021         mtx_lock(&shm_timestamp_lock);
1022         error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
1023             accmode, ucred);
1024         mtx_unlock(&shm_timestamp_lock);
1025         return (error);
1026 }
1027
1028 static void
1029 shm_init(void *arg)
1030 {
1031         char name[32];
1032         int i;
1033
1034         mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF);
1035         sx_init(&shm_dict_lock, "shm dictionary");
1036         shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash);
1037         new_unrhdr64(&shm_ino_unr, 1);
1038         shm_dev_ino = devfs_alloc_cdp_inode();
1039         KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized"));
1040         shmfd_pager_type = vm_pager_alloc_dyn_type(&shm_swap_pager_ops,
1041             OBJT_SWAP);
1042         MPASS(shmfd_pager_type != -1);
1043
1044         for (i = 1; i < MAXPAGESIZES; i++) {
1045                 if (pagesizes[i] == 0)
1046                         break;
1047 #define M       (1024 * 1024)
1048 #define G       (1024 * M)
1049                 if (pagesizes[i] >= G)
1050                         snprintf(name, sizeof(name), "%luG", pagesizes[i] / G);
1051                 else if (pagesizes[i] >= M)
1052                         snprintf(name, sizeof(name), "%luM", pagesizes[i] / M);
1053                 else
1054                         snprintf(name, sizeof(name), "%lu", pagesizes[i]);
1055 #undef G
1056 #undef M
1057                 SYSCTL_ADD_ULONG(NULL, SYSCTL_STATIC_CHILDREN(_vm_largepages),
1058                     OID_AUTO, name, CTLFLAG_RD, &count_largepages[i],
1059                     "number of non-transient largepages allocated");
1060         }
1061 }
1062 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL);
1063
1064 /*
1065  * Remove all shared memory objects that belong to a prison.
1066  */
1067 void
1068 shm_remove_prison(struct prison *pr)
1069 {
1070         struct shm_mapping *shmm, *tshmm;
1071         u_long i;
1072
1073         sx_xlock(&shm_dict_lock);
1074         for (i = 0; i < shm_hash + 1; i++) {
1075                 LIST_FOREACH_SAFE(shmm, &shm_dictionary[i], sm_link, tshmm) {
1076                         if (shmm->sm_shmfd->shm_object->cred &&
1077                             shmm->sm_shmfd->shm_object->cred->cr_prison == pr)
1078                                 shm_doremove(shmm);
1079                 }
1080         }
1081         sx_xunlock(&shm_dict_lock);
1082 }
1083
1084 /*
1085  * Dictionary management.  We maintain an in-kernel dictionary to map
1086  * paths to shmfd objects.  We use the FNV hash on the path to store
1087  * the mappings in a hash table.
1088  */
1089 static struct shmfd *
1090 shm_lookup(char *path, Fnv32_t fnv)
1091 {
1092         struct shm_mapping *map;
1093
1094         LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
1095                 if (map->sm_fnv != fnv)
1096                         continue;
1097                 if (strcmp(map->sm_path, path) == 0)
1098                         return (map->sm_shmfd);
1099         }
1100
1101         return (NULL);
1102 }
1103
1104 static void
1105 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd)
1106 {
1107         struct shm_mapping *map;
1108
1109         map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK);
1110         map->sm_path = path;
1111         map->sm_fnv = fnv;
1112         map->sm_shmfd = shm_hold(shmfd);
1113         shmfd->shm_path = path;
1114         LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link);
1115 }
1116
1117 static int
1118 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred)
1119 {
1120         struct shm_mapping *map;
1121         int error;
1122
1123         LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
1124                 if (map->sm_fnv != fnv)
1125                         continue;
1126                 if (strcmp(map->sm_path, path) == 0) {
1127 #ifdef MAC
1128                         error = mac_posixshm_check_unlink(ucred, map->sm_shmfd);
1129                         if (error)
1130                                 return (error);
1131 #endif
1132                         error = shm_access(map->sm_shmfd, ucred,
1133                             FREAD | FWRITE);
1134                         if (error)
1135                                 return (error);
1136                         shm_doremove(map);
1137                         return (0);
1138                 }
1139         }
1140
1141         return (ENOENT);
1142 }
1143
1144 static void
1145 shm_doremove(struct shm_mapping *map)
1146 {
1147         map->sm_shmfd->shm_path = NULL;
1148         LIST_REMOVE(map, sm_link);
1149         shm_drop(map->sm_shmfd);
1150         free(map->sm_path, M_SHMFD);
1151         free(map, M_SHMFD);
1152 }
1153
1154 int
1155 kern_shm_open2(struct thread *td, const char *userpath, int flags, mode_t mode,
1156     int shmflags, struct filecaps *fcaps, const char *name __unused)
1157 {
1158         struct pwddesc *pdp;
1159         struct shmfd *shmfd;
1160         struct file *fp;
1161         char *path;
1162         void *rl_cookie;
1163         Fnv32_t fnv;
1164         mode_t cmode;
1165         int error, fd, initial_seals;
1166         bool largepage;
1167
1168         if ((shmflags & ~(SHM_ALLOW_SEALING | SHM_GROW_ON_WRITE |
1169             SHM_LARGEPAGE)) != 0)
1170                 return (EINVAL);
1171
1172         initial_seals = F_SEAL_SEAL;
1173         if ((shmflags & SHM_ALLOW_SEALING) != 0)
1174                 initial_seals &= ~F_SEAL_SEAL;
1175
1176 #ifdef CAPABILITY_MODE
1177         /*
1178          * shm_open(2) is only allowed for anonymous objects.
1179          */
1180         if (IN_CAPABILITY_MODE(td) && (userpath != SHM_ANON))
1181                 return (ECAPMODE);
1182 #endif
1183
1184         AUDIT_ARG_FFLAGS(flags);
1185         AUDIT_ARG_MODE(mode);
1186
1187         if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR)
1188                 return (EINVAL);
1189
1190         if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0)
1191                 return (EINVAL);
1192
1193         largepage = (shmflags & SHM_LARGEPAGE) != 0;
1194         if (largepage && !PMAP_HAS_LARGEPAGES)
1195                 return (ENOTTY);
1196
1197         /*
1198          * Currently only F_SEAL_SEAL may be set when creating or opening shmfd.
1199          * If the decision is made later to allow additional seals, care must be
1200          * taken below to ensure that the seals are properly set if the shmfd
1201          * already existed -- this currently assumes that only F_SEAL_SEAL can
1202          * be set and doesn't take further precautions to ensure the validity of
1203          * the seals being added with respect to current mappings.
1204          */
1205         if ((initial_seals & ~F_SEAL_SEAL) != 0)
1206                 return (EINVAL);
1207
1208         pdp = td->td_proc->p_pd;
1209         cmode = (mode & ~pdp->pd_cmask) & ACCESSPERMS;
1210
1211         /*
1212          * shm_open(2) created shm should always have O_CLOEXEC set, as mandated
1213          * by POSIX.  We allow it to be unset here so that an in-kernel
1214          * interface may be written as a thin layer around shm, optionally not
1215          * setting CLOEXEC.  For shm_open(2), O_CLOEXEC is set unconditionally
1216          * in sys_shm_open() to keep this implementation compliant.
1217          */
1218         error = falloc_caps(td, &fp, &fd, flags & O_CLOEXEC, fcaps);
1219         if (error)
1220                 return (error);
1221
1222         /* A SHM_ANON path pointer creates an anonymous object. */
1223         if (userpath == SHM_ANON) {
1224                 /* A read-only anonymous object is pointless. */
1225                 if ((flags & O_ACCMODE) == O_RDONLY) {
1226                         fdclose(td, fp, fd);
1227                         fdrop(fp, td);
1228                         return (EINVAL);
1229                 }
1230                 shmfd = shm_alloc(td->td_ucred, cmode, largepage);
1231                 shmfd->shm_seals = initial_seals;
1232                 shmfd->shm_flags = shmflags;
1233         } else {
1234                 error = shm_copyin_path(td, userpath, &path);
1235                 if (error != 0) {
1236                         fdclose(td, fp, fd);
1237                         fdrop(fp, td);
1238                         return (error);
1239                 }
1240
1241                 AUDIT_ARG_UPATH1_CANON(path);
1242                 fnv = fnv_32_str(path, FNV1_32_INIT);
1243                 sx_xlock(&shm_dict_lock);
1244                 shmfd = shm_lookup(path, fnv);
1245                 if (shmfd == NULL) {
1246                         /* Object does not yet exist, create it if requested. */
1247                         if (flags & O_CREAT) {
1248 #ifdef MAC
1249                                 error = mac_posixshm_check_create(td->td_ucred,
1250                                     path);
1251                                 if (error == 0) {
1252 #endif
1253                                         shmfd = shm_alloc(td->td_ucred, cmode,
1254                                             largepage);
1255                                         shmfd->shm_seals = initial_seals;
1256                                         shmfd->shm_flags = shmflags;
1257                                         shm_insert(path, fnv, shmfd);
1258 #ifdef MAC
1259                                 }
1260 #endif
1261                         } else {
1262                                 free(path, M_SHMFD);
1263                                 error = ENOENT;
1264                         }
1265                 } else {
1266                         rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX);
1267
1268                         /*
1269                          * kern_shm_open() likely shouldn't ever error out on
1270                          * trying to set a seal that already exists, unlike
1271                          * F_ADD_SEALS.  This would break terribly as
1272                          * shm_open(2) actually sets F_SEAL_SEAL to maintain
1273                          * historical behavior where the underlying file could
1274                          * not be sealed.
1275                          */
1276                         initial_seals &= ~shmfd->shm_seals;
1277
1278                         /*
1279                          * Object already exists, obtain a new
1280                          * reference if requested and permitted.
1281                          */
1282                         free(path, M_SHMFD);
1283
1284                         /*
1285                          * initial_seals can't set additional seals if we've
1286                          * already been set F_SEAL_SEAL.  If F_SEAL_SEAL is set,
1287                          * then we've already removed that one from
1288                          * initial_seals.  This is currently redundant as we
1289                          * only allow setting F_SEAL_SEAL at creation time, but
1290                          * it's cheap to check and decreases the effort required
1291                          * to allow additional seals.
1292                          */
1293                         if ((shmfd->shm_seals & F_SEAL_SEAL) != 0 &&
1294                             initial_seals != 0)
1295                                 error = EPERM;
1296                         else if ((flags & (O_CREAT | O_EXCL)) ==
1297                             (O_CREAT | O_EXCL))
1298                                 error = EEXIST;
1299                         else if (shmflags != 0 && shmflags != shmfd->shm_flags)
1300                                 error = EINVAL;
1301                         else {
1302 #ifdef MAC
1303                                 error = mac_posixshm_check_open(td->td_ucred,
1304                                     shmfd, FFLAGS(flags & O_ACCMODE));
1305                                 if (error == 0)
1306 #endif
1307                                 error = shm_access(shmfd, td->td_ucred,
1308                                     FFLAGS(flags & O_ACCMODE));
1309                         }
1310
1311                         /*
1312                          * Truncate the file back to zero length if
1313                          * O_TRUNC was specified and the object was
1314                          * opened with read/write.
1315                          */
1316                         if (error == 0 &&
1317                             (flags & (O_ACCMODE | O_TRUNC)) ==
1318                             (O_RDWR | O_TRUNC)) {
1319                                 VM_OBJECT_WLOCK(shmfd->shm_object);
1320 #ifdef MAC
1321                                 error = mac_posixshm_check_truncate(
1322                                         td->td_ucred, fp->f_cred, shmfd);
1323                                 if (error == 0)
1324 #endif
1325                                         error = shm_dotruncate_locked(shmfd, 0,
1326                                             rl_cookie);
1327                                 VM_OBJECT_WUNLOCK(shmfd->shm_object);
1328                         }
1329                         if (error == 0) {
1330                                 /*
1331                                  * Currently we only allow F_SEAL_SEAL to be
1332                                  * set initially.  As noted above, this would
1333                                  * need to be reworked should that change.
1334                                  */
1335                                 shmfd->shm_seals |= initial_seals;
1336                                 shm_hold(shmfd);
1337                         }
1338                         shm_rangelock_unlock(shmfd, rl_cookie);
1339                 }
1340                 sx_xunlock(&shm_dict_lock);
1341
1342                 if (error) {
1343                         fdclose(td, fp, fd);
1344                         fdrop(fp, td);
1345                         return (error);
1346                 }
1347         }
1348
1349         finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops);
1350
1351         td->td_retval[0] = fd;
1352         fdrop(fp, td);
1353
1354         return (0);
1355 }
1356
1357 /* System calls. */
1358 #ifdef COMPAT_FREEBSD12
1359 int
1360 freebsd12_shm_open(struct thread *td, struct freebsd12_shm_open_args *uap)
1361 {
1362
1363         return (kern_shm_open(td, uap->path, uap->flags | O_CLOEXEC,
1364             uap->mode, NULL));
1365 }
1366 #endif
1367
1368 int
1369 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap)
1370 {
1371         char *path;
1372         Fnv32_t fnv;
1373         int error;
1374
1375         error = shm_copyin_path(td, uap->path, &path);
1376         if (error != 0)
1377                 return (error);
1378
1379         AUDIT_ARG_UPATH1_CANON(path);
1380         fnv = fnv_32_str(path, FNV1_32_INIT);
1381         sx_xlock(&shm_dict_lock);
1382         error = shm_remove(path, fnv, td->td_ucred);
1383         sx_xunlock(&shm_dict_lock);
1384         free(path, M_SHMFD);
1385
1386         return (error);
1387 }
1388
1389 int
1390 sys_shm_rename(struct thread *td, struct shm_rename_args *uap)
1391 {
1392         char *path_from = NULL, *path_to = NULL;
1393         Fnv32_t fnv_from, fnv_to;
1394         struct shmfd *fd_from;
1395         struct shmfd *fd_to;
1396         int error;
1397         int flags;
1398
1399         flags = uap->flags;
1400         AUDIT_ARG_FFLAGS(flags);
1401
1402         /*
1403          * Make sure the user passed only valid flags.
1404          * If you add a new flag, please add a new term here.
1405          */
1406         if ((flags & ~(
1407             SHM_RENAME_NOREPLACE |
1408             SHM_RENAME_EXCHANGE
1409             )) != 0) {
1410                 error = EINVAL;
1411                 goto out;
1412         }
1413
1414         /*
1415          * EXCHANGE and NOREPLACE don't quite make sense together. Let's
1416          * force the user to choose one or the other.
1417          */
1418         if ((flags & SHM_RENAME_NOREPLACE) != 0 &&
1419             (flags & SHM_RENAME_EXCHANGE) != 0) {
1420                 error = EINVAL;
1421                 goto out;
1422         }
1423
1424         /* Renaming to or from anonymous makes no sense */
1425         if (uap->path_from == SHM_ANON || uap->path_to == SHM_ANON) {
1426                 error = EINVAL;
1427                 goto out;
1428         }
1429
1430         error = shm_copyin_path(td, uap->path_from, &path_from);
1431         if (error != 0)
1432                 goto out;
1433
1434         error = shm_copyin_path(td, uap->path_to, &path_to);
1435         if (error != 0)
1436                 goto out;
1437
1438         AUDIT_ARG_UPATH1_CANON(path_from);
1439         AUDIT_ARG_UPATH2_CANON(path_to);
1440
1441         /* Rename with from/to equal is a no-op */
1442         if (strcmp(path_from, path_to) == 0)
1443                 goto out;
1444
1445         fnv_from = fnv_32_str(path_from, FNV1_32_INIT);
1446         fnv_to = fnv_32_str(path_to, FNV1_32_INIT);
1447
1448         sx_xlock(&shm_dict_lock);
1449
1450         fd_from = shm_lookup(path_from, fnv_from);
1451         if (fd_from == NULL) {
1452                 error = ENOENT;
1453                 goto out_locked;
1454         }
1455
1456         fd_to = shm_lookup(path_to, fnv_to);
1457         if ((flags & SHM_RENAME_NOREPLACE) != 0 && fd_to != NULL) {
1458                 error = EEXIST;
1459                 goto out_locked;
1460         }
1461
1462         /*
1463          * Unconditionally prevents shm_remove from invalidating the 'from'
1464          * shm's state.
1465          */
1466         shm_hold(fd_from);
1467         error = shm_remove(path_from, fnv_from, td->td_ucred);
1468
1469         /*
1470          * One of my assumptions failed if ENOENT (e.g. locking didn't
1471          * protect us)
1472          */
1473         KASSERT(error != ENOENT, ("Our shm disappeared during shm_rename: %s",
1474             path_from));
1475         if (error != 0) {
1476                 shm_drop(fd_from);
1477                 goto out_locked;
1478         }
1479
1480         /*
1481          * If we are exchanging, we need to ensure the shm_remove below
1482          * doesn't invalidate the dest shm's state.
1483          */
1484         if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL)
1485                 shm_hold(fd_to);
1486
1487         /*
1488          * NOTE: if path_to is not already in the hash, c'est la vie;
1489          * it simply means we have nothing already at path_to to unlink.
1490          * That is the ENOENT case.
1491          *
1492          * If we somehow don't have access to unlink this guy, but
1493          * did for the shm at path_from, then relink the shm to path_from
1494          * and abort with EACCES.
1495          *
1496          * All other errors: that is weird; let's relink and abort the
1497          * operation.
1498          */
1499         error = shm_remove(path_to, fnv_to, td->td_ucred);
1500         if (error != 0 && error != ENOENT) {
1501                 shm_insert(path_from, fnv_from, fd_from);
1502                 shm_drop(fd_from);
1503                 /* Don't free path_from now, since the hash references it */
1504                 path_from = NULL;
1505                 goto out_locked;
1506         }
1507
1508         error = 0;
1509
1510         shm_insert(path_to, fnv_to, fd_from);
1511
1512         /* Don't free path_to now, since the hash references it */
1513         path_to = NULL;
1514
1515         /* We kept a ref when we removed, and incremented again in insert */
1516         shm_drop(fd_from);
1517         KASSERT(fd_from->shm_refs > 0, ("Expected >0 refs; got: %d\n",
1518             fd_from->shm_refs));
1519
1520         if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) {
1521                 shm_insert(path_from, fnv_from, fd_to);
1522                 path_from = NULL;
1523                 shm_drop(fd_to);
1524                 KASSERT(fd_to->shm_refs > 0, ("Expected >0 refs; got: %d\n",
1525                     fd_to->shm_refs));
1526         }
1527
1528 out_locked:
1529         sx_xunlock(&shm_dict_lock);
1530
1531 out:
1532         free(path_from, M_SHMFD);
1533         free(path_to, M_SHMFD);
1534         return (error);
1535 }
1536
1537 static int
1538 shm_mmap_large(struct shmfd *shmfd, vm_map_t map, vm_offset_t *addr,
1539     vm_size_t size, vm_prot_t prot, vm_prot_t max_prot, int flags,
1540     vm_ooffset_t foff, struct thread *td)
1541 {
1542         struct vmspace *vms;
1543         vm_map_entry_t next_entry, prev_entry;
1544         vm_offset_t align, mask, maxaddr;
1545         int docow, error, rv, try;
1546         bool curmap;
1547
1548         if (shmfd->shm_lp_psind == 0)
1549                 return (EINVAL);
1550
1551         /* MAP_PRIVATE is disabled */
1552         if ((flags & ~(MAP_SHARED | MAP_FIXED | MAP_EXCL |
1553             MAP_NOCORE | MAP_32BIT | MAP_ALIGNMENT_MASK)) != 0)
1554                 return (EINVAL);
1555
1556         vms = td->td_proc->p_vmspace;
1557         curmap = map == &vms->vm_map;
1558         if (curmap) {
1559                 error = kern_mmap_racct_check(td, map, size);
1560                 if (error != 0)
1561                         return (error);
1562         }
1563
1564         docow = shmfd->shm_lp_psind << MAP_SPLIT_BOUNDARY_SHIFT;
1565         docow |= MAP_INHERIT_SHARE;
1566         if ((flags & MAP_NOCORE) != 0)
1567                 docow |= MAP_DISABLE_COREDUMP;
1568
1569         mask = pagesizes[shmfd->shm_lp_psind] - 1;
1570         if ((foff & mask) != 0)
1571                 return (EINVAL);
1572         maxaddr = vm_map_max(map);
1573         if ((flags & MAP_32BIT) != 0 && maxaddr > MAP_32BIT_MAX_ADDR)
1574                 maxaddr = MAP_32BIT_MAX_ADDR;
1575         if (size == 0 || (size & mask) != 0 ||
1576             (*addr != 0 && ((*addr & mask) != 0 ||
1577             *addr + size < *addr || *addr + size > maxaddr)))
1578                 return (EINVAL);
1579
1580         align = flags & MAP_ALIGNMENT_MASK;
1581         if (align == 0) {
1582                 align = pagesizes[shmfd->shm_lp_psind];
1583         } else if (align == MAP_ALIGNED_SUPER) {
1584                 if (shmfd->shm_lp_psind != 1)
1585                         return (EINVAL);
1586                 align = pagesizes[1];
1587         } else {
1588                 align >>= MAP_ALIGNMENT_SHIFT;
1589                 align = 1ULL << align;
1590                 /* Also handles overflow. */
1591                 if (align < pagesizes[shmfd->shm_lp_psind])
1592                         return (EINVAL);
1593         }
1594
1595         vm_map_lock(map);
1596         if ((flags & MAP_FIXED) == 0) {
1597                 try = 1;
1598                 if (curmap && (*addr == 0 ||
1599                     (*addr >= round_page((vm_offset_t)vms->vm_taddr) &&
1600                     *addr < round_page((vm_offset_t)vms->vm_daddr +
1601                     lim_max(td, RLIMIT_DATA))))) {
1602                         *addr = roundup2((vm_offset_t)vms->vm_daddr +
1603                             lim_max(td, RLIMIT_DATA),
1604                             pagesizes[shmfd->shm_lp_psind]);
1605                 }
1606 again:
1607                 rv = vm_map_find_aligned(map, addr, size, maxaddr, align);
1608                 if (rv != KERN_SUCCESS) {
1609                         if (try == 1) {
1610                                 try = 2;
1611                                 *addr = vm_map_min(map);
1612                                 if ((*addr & mask) != 0)
1613                                         *addr = (*addr + mask) & mask;
1614                                 goto again;
1615                         }
1616                         goto fail1;
1617                 }
1618         } else if ((flags & MAP_EXCL) == 0) {
1619                 rv = vm_map_delete(map, *addr, *addr + size);
1620                 if (rv != KERN_SUCCESS)
1621                         goto fail1;
1622         } else {
1623                 error = ENOSPC;
1624                 if (vm_map_lookup_entry(map, *addr, &prev_entry))
1625                         goto fail;
1626                 next_entry = vm_map_entry_succ(prev_entry);
1627                 if (next_entry->start < *addr + size)
1628                         goto fail;
1629         }
1630
1631         rv = vm_map_insert(map, shmfd->shm_object, foff, *addr, *addr + size,
1632             prot, max_prot, docow);
1633 fail1:
1634         error = vm_mmap_to_errno(rv);
1635 fail:
1636         vm_map_unlock(map);
1637         return (error);
1638 }
1639
1640 static int
1641 shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize,
1642     vm_prot_t prot, vm_prot_t cap_maxprot, int flags,
1643     vm_ooffset_t foff, struct thread *td)
1644 {
1645         struct shmfd *shmfd;
1646         vm_prot_t maxprot;
1647         int error;
1648         bool writecnt;
1649         void *rl_cookie;
1650
1651         shmfd = fp->f_data;
1652         maxprot = VM_PROT_NONE;
1653
1654         rl_cookie = shm_rangelock_rlock(shmfd, 0, objsize);
1655         /* FREAD should always be set. */
1656         if ((fp->f_flag & FREAD) != 0)
1657                 maxprot |= VM_PROT_EXECUTE | VM_PROT_READ;
1658
1659         /*
1660          * If FWRITE's set, we can allow VM_PROT_WRITE unless it's a shared
1661          * mapping with a write seal applied.  Private mappings are always
1662          * writeable.
1663          */
1664         if ((flags & MAP_SHARED) == 0) {
1665                 cap_maxprot |= VM_PROT_WRITE;
1666                 maxprot |= VM_PROT_WRITE;
1667                 writecnt = false;
1668         } else {
1669                 if ((fp->f_flag & FWRITE) != 0 &&
1670                     (shmfd->shm_seals & F_SEAL_WRITE) == 0)
1671                         maxprot |= VM_PROT_WRITE;
1672
1673                 /*
1674                  * Any mappings from a writable descriptor may be upgraded to
1675                  * VM_PROT_WRITE with mprotect(2), unless a write-seal was
1676                  * applied between the open and subsequent mmap(2).  We want to
1677                  * reject application of a write seal as long as any such
1678                  * mapping exists so that the seal cannot be trivially bypassed.
1679                  */
1680                 writecnt = (maxprot & VM_PROT_WRITE) != 0;
1681                 if (!writecnt && (prot & VM_PROT_WRITE) != 0) {
1682                         error = EACCES;
1683                         goto out;
1684                 }
1685         }
1686         maxprot &= cap_maxprot;
1687
1688         /* See comment in vn_mmap(). */
1689         if (
1690 #ifdef _LP64
1691             objsize > OFF_MAX ||
1692 #endif
1693             foff > OFF_MAX - objsize) {
1694                 error = EINVAL;
1695                 goto out;
1696         }
1697
1698 #ifdef MAC
1699         error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags);
1700         if (error != 0)
1701                 goto out;
1702 #endif
1703
1704         mtx_lock(&shm_timestamp_lock);
1705         vfs_timestamp(&shmfd->shm_atime);
1706         mtx_unlock(&shm_timestamp_lock);
1707         vm_object_reference(shmfd->shm_object);
1708
1709         if (shm_largepage(shmfd)) {
1710                 writecnt = false;
1711                 error = shm_mmap_large(shmfd, map, addr, objsize, prot,
1712                     maxprot, flags, foff, td);
1713         } else {
1714                 if (writecnt) {
1715                         vm_pager_update_writecount(shmfd->shm_object, 0,
1716                             objsize);
1717                 }
1718                 error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags,
1719                     shmfd->shm_object, foff, writecnt, td);
1720         }
1721         if (error != 0) {
1722                 if (writecnt)
1723                         vm_pager_release_writecount(shmfd->shm_object, 0,
1724                             objsize);
1725                 vm_object_deallocate(shmfd->shm_object);
1726         }
1727 out:
1728         shm_rangelock_unlock(shmfd, rl_cookie);
1729         return (error);
1730 }
1731
1732 static int
1733 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
1734     struct thread *td)
1735 {
1736         struct shmfd *shmfd;
1737         int error;
1738
1739         error = 0;
1740         shmfd = fp->f_data;
1741         mtx_lock(&shm_timestamp_lock);
1742         /*
1743          * SUSv4 says that x bits of permission need not be affected.
1744          * Be consistent with our shm_open there.
1745          */
1746 #ifdef MAC
1747         error = mac_posixshm_check_setmode(active_cred, shmfd, mode);
1748         if (error != 0)
1749                 goto out;
1750 #endif
1751         error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
1752             VADMIN, active_cred);
1753         if (error != 0)
1754                 goto out;
1755         shmfd->shm_mode = mode & ACCESSPERMS;
1756 out:
1757         mtx_unlock(&shm_timestamp_lock);
1758         return (error);
1759 }
1760
1761 static int
1762 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
1763     struct thread *td)
1764 {
1765         struct shmfd *shmfd;
1766         int error;
1767
1768         error = 0;
1769         shmfd = fp->f_data;
1770         mtx_lock(&shm_timestamp_lock);
1771 #ifdef MAC
1772         error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid);
1773         if (error != 0)
1774                 goto out;
1775 #endif
1776         if (uid == (uid_t)-1)
1777                 uid = shmfd->shm_uid;
1778         if (gid == (gid_t)-1)
1779                  gid = shmfd->shm_gid;
1780         if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) ||
1781             (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) &&
1782             (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN)))
1783                 goto out;
1784         shmfd->shm_uid = uid;
1785         shmfd->shm_gid = gid;
1786 out:
1787         mtx_unlock(&shm_timestamp_lock);
1788         return (error);
1789 }
1790
1791 /*
1792  * Helper routines to allow the backing object of a shared memory file
1793  * descriptor to be mapped in the kernel.
1794  */
1795 int
1796 shm_map(struct file *fp, size_t size, off_t offset, void **memp)
1797 {
1798         struct shmfd *shmfd;
1799         vm_offset_t kva, ofs;
1800         vm_object_t obj;
1801         int rv;
1802
1803         if (fp->f_type != DTYPE_SHM)
1804                 return (EINVAL);
1805         shmfd = fp->f_data;
1806         obj = shmfd->shm_object;
1807         VM_OBJECT_WLOCK(obj);
1808         /*
1809          * XXXRW: This validation is probably insufficient, and subject to
1810          * sign errors.  It should be fixed.
1811          */
1812         if (offset >= shmfd->shm_size ||
1813             offset + size > round_page(shmfd->shm_size)) {
1814                 VM_OBJECT_WUNLOCK(obj);
1815                 return (EINVAL);
1816         }
1817
1818         shmfd->shm_kmappings++;
1819         vm_object_reference_locked(obj);
1820         VM_OBJECT_WUNLOCK(obj);
1821
1822         /* Map the object into the kernel_map and wire it. */
1823         kva = vm_map_min(kernel_map);
1824         ofs = offset & PAGE_MASK;
1825         offset = trunc_page(offset);
1826         size = round_page(size + ofs);
1827         rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0,
1828             VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE,
1829             VM_PROT_READ | VM_PROT_WRITE, 0);
1830         if (rv == KERN_SUCCESS) {
1831                 rv = vm_map_wire(kernel_map, kva, kva + size,
1832                     VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
1833                 if (rv == KERN_SUCCESS) {
1834                         *memp = (void *)(kva + ofs);
1835                         return (0);
1836                 }
1837                 vm_map_remove(kernel_map, kva, kva + size);
1838         } else
1839                 vm_object_deallocate(obj);
1840
1841         /* On failure, drop our mapping reference. */
1842         VM_OBJECT_WLOCK(obj);
1843         shmfd->shm_kmappings--;
1844         VM_OBJECT_WUNLOCK(obj);
1845
1846         return (vm_mmap_to_errno(rv));
1847 }
1848
1849 /*
1850  * We require the caller to unmap the entire entry.  This allows us to
1851  * safely decrement shm_kmappings when a mapping is removed.
1852  */
1853 int
1854 shm_unmap(struct file *fp, void *mem, size_t size)
1855 {
1856         struct shmfd *shmfd;
1857         vm_map_entry_t entry;
1858         vm_offset_t kva, ofs;
1859         vm_object_t obj;
1860         vm_pindex_t pindex;
1861         vm_prot_t prot;
1862         boolean_t wired;
1863         vm_map_t map;
1864         int rv;
1865
1866         if (fp->f_type != DTYPE_SHM)
1867                 return (EINVAL);
1868         shmfd = fp->f_data;
1869         kva = (vm_offset_t)mem;
1870         ofs = kva & PAGE_MASK;
1871         kva = trunc_page(kva);
1872         size = round_page(size + ofs);
1873         map = kernel_map;
1874         rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry,
1875             &obj, &pindex, &prot, &wired);
1876         if (rv != KERN_SUCCESS)
1877                 return (EINVAL);
1878         if (entry->start != kva || entry->end != kva + size) {
1879                 vm_map_lookup_done(map, entry);
1880                 return (EINVAL);
1881         }
1882         vm_map_lookup_done(map, entry);
1883         if (obj != shmfd->shm_object)
1884                 return (EINVAL);
1885         vm_map_remove(map, kva, kva + size);
1886         VM_OBJECT_WLOCK(obj);
1887         KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped"));
1888         shmfd->shm_kmappings--;
1889         VM_OBJECT_WUNLOCK(obj);
1890         return (0);
1891 }
1892
1893 static int
1894 shm_fill_kinfo_locked(struct shmfd *shmfd, struct kinfo_file *kif, bool list)
1895 {
1896         const char *path, *pr_path;
1897         size_t pr_pathlen;
1898         bool visible;
1899
1900         sx_assert(&shm_dict_lock, SA_LOCKED);
1901         kif->kf_type = KF_TYPE_SHM;
1902         kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode;
1903         kif->kf_un.kf_file.kf_file_size = shmfd->shm_size;
1904         if (shmfd->shm_path != NULL) {
1905                 if (shmfd->shm_path != NULL) {
1906                         path = shmfd->shm_path;
1907                         pr_path = curthread->td_ucred->cr_prison->pr_path;
1908                         if (strcmp(pr_path, "/") != 0) {
1909                                 /* Return the jail-rooted pathname. */
1910                                 pr_pathlen = strlen(pr_path);
1911                                 visible = strncmp(path, pr_path, pr_pathlen)
1912                                     == 0 && path[pr_pathlen] == '/';
1913                                 if (list && !visible)
1914                                         return (EPERM);
1915                                 if (visible)
1916                                         path += pr_pathlen;
1917                         }
1918                         strlcpy(kif->kf_path, path, sizeof(kif->kf_path));
1919                 }
1920         }
1921         return (0);
1922 }
1923
1924 static int
1925 shm_fill_kinfo(struct file *fp, struct kinfo_file *kif,
1926     struct filedesc *fdp __unused)
1927 {
1928         int res;
1929
1930         sx_slock(&shm_dict_lock);
1931         res = shm_fill_kinfo_locked(fp->f_data, kif, false);
1932         sx_sunlock(&shm_dict_lock);
1933         return (res);
1934 }
1935
1936 static int
1937 shm_add_seals(struct file *fp, int seals)
1938 {
1939         struct shmfd *shmfd;
1940         void *rl_cookie;
1941         vm_ooffset_t writemappings;
1942         int error, nseals;
1943
1944         error = 0;
1945         shmfd = fp->f_data;
1946         rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX);
1947
1948         /* Even already-set seals should result in EPERM. */
1949         if ((shmfd->shm_seals & F_SEAL_SEAL) != 0) {
1950                 error = EPERM;
1951                 goto out;
1952         }
1953         nseals = seals & ~shmfd->shm_seals;
1954         if ((nseals & F_SEAL_WRITE) != 0) {
1955                 if (shm_largepage(shmfd)) {
1956                         error = ENOTSUP;
1957                         goto out;
1958                 }
1959
1960                 /*
1961                  * The rangelock above prevents writable mappings from being
1962                  * added after we've started applying seals.  The RLOCK here
1963                  * is to avoid torn reads on ILP32 arches as unmapping/reducing
1964                  * writemappings will be done without a rangelock.
1965                  */
1966                 VM_OBJECT_RLOCK(shmfd->shm_object);
1967                 writemappings = shmfd->shm_object->un_pager.swp.writemappings;
1968                 VM_OBJECT_RUNLOCK(shmfd->shm_object);
1969                 /* kmappings are also writable */
1970                 if (writemappings > 0) {
1971                         error = EBUSY;
1972                         goto out;
1973                 }
1974         }
1975         shmfd->shm_seals |= nseals;
1976 out:
1977         shm_rangelock_unlock(shmfd, rl_cookie);
1978         return (error);
1979 }
1980
1981 static int
1982 shm_get_seals(struct file *fp, int *seals)
1983 {
1984         struct shmfd *shmfd;
1985
1986         shmfd = fp->f_data;
1987         *seals = shmfd->shm_seals;
1988         return (0);
1989 }
1990
1991 static int
1992 shm_deallocate(struct shmfd *shmfd, off_t *offset, off_t *length, int flags)
1993 {
1994         vm_object_t object;
1995         vm_pindex_t pistart, pi, piend;
1996         vm_ooffset_t off, len;
1997         int startofs, endofs, end;
1998         int error;
1999
2000         off = *offset;
2001         len = *length;
2002         KASSERT(off + len <= (vm_ooffset_t)OFF_MAX, ("off + len overflows"));
2003         if (off + len > shmfd->shm_size)
2004                 len = shmfd->shm_size - off;
2005         object = shmfd->shm_object;
2006         startofs = off & PAGE_MASK;
2007         endofs = (off + len) & PAGE_MASK;
2008         pistart = OFF_TO_IDX(off);
2009         piend = OFF_TO_IDX(off + len);
2010         pi = OFF_TO_IDX(off + PAGE_MASK);
2011         error = 0;
2012
2013         /* Handle the case when offset is on or beyond shm size. */
2014         if ((off_t)len <= 0) {
2015                 *length = 0;
2016                 return (0);
2017         }
2018
2019         VM_OBJECT_WLOCK(object);
2020
2021         if (startofs != 0) {
2022                 end = pistart != piend ? PAGE_SIZE : endofs;
2023                 error = shm_partial_page_invalidate(object, pistart, startofs,
2024                     end);
2025                 if (error)
2026                         goto out;
2027                 off += end - startofs;
2028                 len -= end - startofs;
2029         }
2030
2031         if (pi < piend) {
2032                 vm_object_page_remove(object, pi, piend, 0);
2033                 off += IDX_TO_OFF(piend - pi);
2034                 len -= IDX_TO_OFF(piend - pi);
2035         }
2036
2037         if (endofs != 0 && pistart != piend) {
2038                 error = shm_partial_page_invalidate(object, piend, 0, endofs);
2039                 if (error)
2040                         goto out;
2041                 off += endofs;
2042                 len -= endofs;
2043         }
2044
2045 out:
2046         VM_OBJECT_WUNLOCK(shmfd->shm_object);
2047         *offset = off;
2048         *length = len;
2049         return (error);
2050 }
2051
2052 static int
2053 shm_fspacectl(struct file *fp, int cmd, off_t *offset, off_t *length, int flags,
2054     struct ucred *active_cred, struct thread *td)
2055 {
2056         void *rl_cookie;
2057         struct shmfd *shmfd;
2058         off_t off, len;
2059         int error;
2060
2061         KASSERT(cmd == SPACECTL_DEALLOC, ("shm_fspacectl: Invalid cmd"));
2062         KASSERT((flags & ~SPACECTL_F_SUPPORTED) == 0,
2063             ("shm_fspacectl: non-zero flags"));
2064         KASSERT(*offset >= 0 && *length > 0 && *length <= OFF_MAX - *offset,
2065             ("shm_fspacectl: offset/length overflow or underflow"));
2066         error = EINVAL;
2067         shmfd = fp->f_data;
2068         off = *offset;
2069         len = *length;
2070
2071         rl_cookie = shm_rangelock_wlock(shmfd, off, off + len);
2072         switch (cmd) {
2073         case SPACECTL_DEALLOC:
2074                 if ((shmfd->shm_seals & F_SEAL_WRITE) != 0) {
2075                         error = EPERM;
2076                         break;
2077                 }
2078                 error = shm_deallocate(shmfd, &off, &len, flags);
2079                 *offset = off;
2080                 *length = len;
2081                 break;
2082         default:
2083                 __assert_unreachable();
2084         }
2085         shm_rangelock_unlock(shmfd, rl_cookie);
2086         return (error);
2087 }
2088
2089
2090 static int
2091 shm_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
2092 {
2093         void *rl_cookie;
2094         struct shmfd *shmfd;
2095         size_t size;
2096         int error;
2097
2098         /* This assumes that the caller already checked for overflow. */
2099         error = 0;
2100         shmfd = fp->f_data;
2101         size = offset + len;
2102
2103         /*
2104          * Just grab the rangelock for the range that we may be attempting to
2105          * grow, rather than blocking read/write for regions we won't be
2106          * touching while this (potential) resize is in progress.  Other
2107          * attempts to resize the shmfd will have to take a write lock from 0 to
2108          * OFF_MAX, so this being potentially beyond the current usable range of
2109          * the shmfd is not necessarily a concern.  If other mechanisms are
2110          * added to grow a shmfd, this may need to be re-evaluated.
2111          */
2112         rl_cookie = shm_rangelock_wlock(shmfd, offset, size);
2113         if (size > shmfd->shm_size)
2114                 error = shm_dotruncate_cookie(shmfd, size, rl_cookie);
2115         shm_rangelock_unlock(shmfd, rl_cookie);
2116         /* Translate to posix_fallocate(2) return value as needed. */
2117         if (error == ENOMEM)
2118                 error = ENOSPC;
2119         return (error);
2120 }
2121
2122 static int
2123 sysctl_posix_shm_list(SYSCTL_HANDLER_ARGS)
2124 {
2125         struct shm_mapping *shmm;
2126         struct sbuf sb;
2127         struct kinfo_file kif;
2128         u_long i;
2129         int error, error2;
2130
2131         sbuf_new_for_sysctl(&sb, NULL, sizeof(struct kinfo_file) * 5, req);
2132         sbuf_clear_flags(&sb, SBUF_INCLUDENUL);
2133         error = 0;
2134         sx_slock(&shm_dict_lock);
2135         for (i = 0; i < shm_hash + 1; i++) {
2136                 LIST_FOREACH(shmm, &shm_dictionary[i], sm_link) {
2137                         error = shm_fill_kinfo_locked(shmm->sm_shmfd,
2138                             &kif, true);
2139                         if (error == EPERM) {
2140                                 error = 0;
2141                                 continue;
2142                         }
2143                         if (error != 0)
2144                                 break;
2145                         pack_kinfo(&kif);
2146                         error = sbuf_bcat(&sb, &kif, kif.kf_structsize) == 0 ?
2147                             0 : ENOMEM;
2148                         if (error != 0)
2149                                 break;
2150                 }
2151         }
2152         sx_sunlock(&shm_dict_lock);
2153         error2 = sbuf_finish(&sb);
2154         sbuf_delete(&sb);
2155         return (error != 0 ? error : error2);
2156 }
2157
2158 SYSCTL_PROC(_kern_ipc, OID_AUTO, posix_shm_list,
2159     CTLFLAG_RD | CTLFLAG_PRISON | CTLFLAG_MPSAFE | CTLTYPE_OPAQUE,
2160     NULL, 0, sysctl_posix_shm_list, "",
2161     "POSIX SHM list");
2162
2163 int
2164 kern_shm_open(struct thread *td, const char *path, int flags, mode_t mode,
2165     struct filecaps *caps)
2166 {
2167
2168         return (kern_shm_open2(td, path, flags, mode, 0, caps, NULL));
2169 }
2170
2171 /*
2172  * This version of the shm_open() interface leaves CLOEXEC behavior up to the
2173  * caller, and libc will enforce it for the traditional shm_open() call.  This
2174  * allows other consumers, like memfd_create(), to opt-in for CLOEXEC.  This
2175  * interface also includes a 'name' argument that is currently unused, but could
2176  * potentially be exported later via some interface for debugging purposes.
2177  * From the kernel's perspective, it is optional.  Individual consumers like
2178  * memfd_create() may require it in order to be compatible with other systems
2179  * implementing the same function.
2180  */
2181 int
2182 sys_shm_open2(struct thread *td, struct shm_open2_args *uap)
2183 {
2184
2185         return (kern_shm_open2(td, uap->path, uap->flags, uap->mode,
2186             uap->shmflags, NULL, uap->name));
2187 }