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/vm_radix.h> 104 #include <vm/swap_pager.h> 105 106 struct shm_mapping { 107 char *sm_path; 108 Fnv32_t sm_fnv; 109 struct shmfd *sm_shmfd; 110 LIST_ENTRY(shm_mapping) sm_link; 111 }; 112 113 static MALLOC_DEFINE(M_SHMFD, "shmfd", "shared memory file descriptor"); 114 static LIST_HEAD(, shm_mapping) *shm_dictionary; 115 static struct sx shm_dict_lock; 116 static struct mtx shm_timestamp_lock; 117 static u_long shm_hash; 118 static struct unrhdr64 shm_ino_unr; 119 static dev_t shm_dev_ino; 120 121 #define SHM_HASH(fnv) (&shm_dictionary[(fnv) & shm_hash]) 122 123 static void shm_init(void *arg); 124 static void shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd); 125 static struct shmfd *shm_lookup(char *path, Fnv32_t fnv); 126 static int shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred); 127 static void shm_doremove(struct shm_mapping *map); 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 static int shm_deallocate(struct shmfd *shmfd, off_t *offset, 135 off_t *length, int flags); 136 137 static fo_rdwr_t shm_read; 138 static fo_rdwr_t shm_write; 139 static fo_truncate_t shm_truncate; 140 static fo_ioctl_t shm_ioctl; 141 static fo_stat_t shm_stat; 142 static fo_close_t shm_close; 143 static fo_chmod_t shm_chmod; 144 static fo_chown_t shm_chown; 145 static fo_seek_t shm_seek; 146 static fo_fill_kinfo_t shm_fill_kinfo; 147 static fo_mmap_t shm_mmap; 148 static fo_get_seals_t shm_get_seals; 149 static fo_add_seals_t shm_add_seals; 150 static fo_fallocate_t shm_fallocate; 151 static fo_fspacectl_t shm_fspacectl; 152 153 /* File descriptor operations. */ 154 const struct fileops shm_ops = { 155 .fo_read = shm_read, 156 .fo_write = shm_write, 157 .fo_truncate = shm_truncate, 158 .fo_ioctl = shm_ioctl, 159 .fo_poll = invfo_poll, 160 .fo_kqfilter = invfo_kqfilter, 161 .fo_stat = shm_stat, 162 .fo_close = shm_close, 163 .fo_chmod = shm_chmod, 164 .fo_chown = shm_chown, 165 .fo_sendfile = vn_sendfile, 166 .fo_seek = shm_seek, 167 .fo_fill_kinfo = shm_fill_kinfo, 168 .fo_mmap = shm_mmap, 169 .fo_get_seals = shm_get_seals, 170 .fo_add_seals = shm_add_seals, 171 .fo_fallocate = shm_fallocate, 172 .fo_fspacectl = shm_fspacectl, 173 .fo_cmp = file_kcmp_generic, 174 .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE, 175 }; 176 177 FEATURE(posix_shm, "POSIX shared memory"); 178 179 static SYSCTL_NODE(_vm, OID_AUTO, largepages, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 180 ""); 181 182 static int largepage_reclaim_tries = 1; 183 SYSCTL_INT(_vm_largepages, OID_AUTO, reclaim_tries, 184 CTLFLAG_RWTUN, &largepage_reclaim_tries, 0, 185 "Number of contig reclaims before giving up for default alloc policy"); 186 187 #define shm_rangelock_unlock(shmfd, cookie) \ 188 rangelock_unlock(&(shmfd)->shm_rl, (cookie)) 189 #define shm_rangelock_rlock(shmfd, start, end) \ 190 rangelock_rlock(&(shmfd)->shm_rl, (start), (end)) 191 #define shm_rangelock_tryrlock(shmfd, start, end) \ 192 rangelock_tryrlock(&(shmfd)->shm_rl, (start), (end)) 193 #define shm_rangelock_wlock(shmfd, start, end) \ 194 rangelock_wlock(&(shmfd)->shm_rl, (start), (end)) 195 196 static int 197 uiomove_object_page(vm_object_t obj, size_t len, struct uio *uio) 198 { 199 struct pctrie_iter pages; 200 vm_page_t m; 201 vm_pindex_t idx; 202 size_t tlen; 203 int error, offset, rv; 204 205 idx = OFF_TO_IDX(uio->uio_offset); 206 offset = uio->uio_offset & PAGE_MASK; 207 tlen = MIN(PAGE_SIZE - offset, len); 208 209 rv = vm_page_grab_valid_unlocked(&m, obj, idx, 210 VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY | VM_ALLOC_NOCREAT); 211 if (rv == VM_PAGER_OK) 212 goto found; 213 214 /* 215 * Read I/O without either a corresponding resident page or swap 216 * page: use zero_region. This is intended to avoid instantiating 217 * pages on read from a sparse region. 218 */ 219 vm_page_iter_init(&pages, obj); 220 VM_OBJECT_WLOCK(obj); 221 m = vm_radix_iter_lookup(&pages, idx); 222 if (uio->uio_rw == UIO_READ && m == NULL && 223 !vm_pager_has_page(obj, idx, NULL, NULL)) { 224 VM_OBJECT_WUNLOCK(obj); 225 return (uiomove(__DECONST(void *, zero_region), tlen, uio)); 226 } 227 228 /* 229 * Although the tmpfs vnode lock is held here, it is 230 * nonetheless safe to sleep waiting for a free page. The 231 * pageout daemon does not need to acquire the tmpfs vnode 232 * lock to page out tobj's pages because tobj is a OBJT_SWAP 233 * type object. 234 */ 235 rv = vm_page_grab_valid_iter(&m, obj, idx, 236 VM_ALLOC_NORMAL | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY, &pages); 237 if (rv != VM_PAGER_OK) { 238 VM_OBJECT_WUNLOCK(obj); 239 if (bootverbose) { 240 printf("uiomove_object: vm_obj %p idx %jd " 241 "pager error %d\n", obj, idx, rv); 242 } 243 return (rv == VM_PAGER_AGAIN ? ENOSPC : EIO); 244 } 245 VM_OBJECT_WUNLOCK(obj); 246 247 found: 248 error = uiomove_fromphys(&m, offset, tlen, uio); 249 if (uio->uio_rw == UIO_WRITE) { 250 /* 251 * Even a failed copy may have changed the page's contents, 252 * because uiomove_fromphys() can copy part of the data before 253 * it fails. 254 */ 255 vm_page_set_dirty(m); 256 } 257 vm_page_activate(m); 258 vm_page_sunbusy(m); 259 260 return (error); 261 } 262 263 int 264 uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio) 265 { 266 ssize_t resid; 267 size_t len; 268 int error; 269 270 error = 0; 271 while ((resid = uio->uio_resid) > 0) { 272 if (obj_size <= uio->uio_offset) 273 break; 274 len = MIN(obj_size - uio->uio_offset, resid); 275 if (len == 0) 276 break; 277 error = uiomove_object_page(obj, len, uio); 278 if (error != 0 || resid == uio->uio_resid) 279 break; 280 } 281 return (error); 282 } 283 284 static u_long count_largepages[MAXPAGESIZES]; 285 286 static int 287 shm_largepage_phys_populate(vm_object_t object, vm_pindex_t pidx, 288 int fault_type, vm_prot_t max_prot, vm_pindex_t *first, vm_pindex_t *last) 289 { 290 vm_page_t m __diagused; 291 int psind; 292 293 psind = object->un_pager.phys.data_val; 294 if (psind == 0 || pidx >= object->size) 295 return (VM_PAGER_FAIL); 296 *first = rounddown2(pidx, pagesizes[psind] / PAGE_SIZE); 297 298 /* 299 * We only busy the first page in the superpage run. It is 300 * useless to busy whole run since we only remove full 301 * superpage, and it takes too long to busy e.g. 512 * 512 == 302 * 262144 pages constituing 1G amd64 superage. 303 */ 304 m = vm_page_grab(object, *first, VM_ALLOC_NORMAL | VM_ALLOC_NOCREAT); 305 MPASS(m != NULL); 306 307 *last = *first + atop(pagesizes[psind]) - 1; 308 return (VM_PAGER_OK); 309 } 310 311 static boolean_t 312 shm_largepage_phys_haspage(vm_object_t object, vm_pindex_t pindex, 313 int *before, int *after) 314 { 315 int psind; 316 317 psind = object->un_pager.phys.data_val; 318 if (psind == 0 || pindex >= object->size) 319 return (FALSE); 320 if (before != NULL) { 321 *before = pindex - rounddown2(pindex, pagesizes[psind] / 322 PAGE_SIZE); 323 } 324 if (after != NULL) { 325 *after = roundup2(pindex, pagesizes[psind] / PAGE_SIZE) - 326 pindex; 327 } 328 return (TRUE); 329 } 330 331 static void 332 shm_largepage_phys_ctor(vm_object_t object, vm_prot_t prot, 333 vm_ooffset_t foff, struct ucred *cred) 334 { 335 object->flags |= OBJ_PG_DTOR; 336 } 337 338 static void 339 shm_largepage_phys_dtor(vm_object_t object) 340 { 341 int psind; 342 343 VM_OBJECT_ASSERT_WLOCKED(object); 344 345 psind = object->un_pager.phys.data_val; 346 if (psind != 0) { 347 struct pctrie_iter pages; 348 vm_page_t m; 349 bool removed __diagused; 350 351 vm_page_iter_init(&pages, object); 352 restart: 353 VM_RADIX_FOREACH(m, &pages) { 354 if (!vm_page_busy_acquire(m, VM_ALLOC_WAITFAIL)) { 355 pctrie_iter_reset(&pages); 356 goto restart; 357 } 358 removed = vm_page_iter_remove(&pages, m); 359 KASSERT(!removed, ("%s: page %p not wired", __func__, m)); 360 vm_page_unwire(m, PQ_NONE); 361 } 362 atomic_subtract_long(&count_largepages[psind], 363 object->size / (pagesizes[psind] / PAGE_SIZE)); 364 } else { 365 KASSERT(object->size == 0, 366 ("largepage phys obj %p not initialized bit size %#jx > 0", 367 object, (uintmax_t)object->size)); 368 } 369 } 370 371 static const struct phys_pager_ops shm_largepage_phys_ops = { 372 .phys_pg_populate = shm_largepage_phys_populate, 373 .phys_pg_haspage = shm_largepage_phys_haspage, 374 .phys_pg_ctor = shm_largepage_phys_ctor, 375 .phys_pg_dtor = shm_largepage_phys_dtor, 376 }; 377 378 bool 379 shm_largepage(struct shmfd *shmfd) 380 { 381 return (shmfd->shm_object->type == OBJT_PHYS); 382 } 383 384 static void 385 shm_pager_freespace(vm_object_t obj, vm_pindex_t start, vm_size_t size) 386 { 387 struct shmfd *shm; 388 vm_size_t c; 389 390 swap_pager_freespace(obj, start, size, &c); 391 if (c == 0) 392 return; 393 394 shm = obj->un_pager.swp.swp_priv; 395 if (shm == NULL) 396 return; 397 KASSERT(shm->shm_pages >= c, 398 ("shm %p pages %jd free %jd", shm, 399 (uintmax_t)shm->shm_pages, (uintmax_t)c)); 400 shm->shm_pages -= c; 401 } 402 403 static void 404 shm_page_inserted(vm_object_t obj, vm_page_t m) 405 { 406 struct shmfd *shm; 407 408 shm = obj->un_pager.swp.swp_priv; 409 if (shm == NULL) 410 return; 411 if (!vm_pager_has_page(obj, m->pindex, NULL, NULL)) 412 shm->shm_pages += 1; 413 } 414 415 static void 416 shm_page_removed(vm_object_t obj, vm_page_t m) 417 { 418 struct shmfd *shm; 419 420 shm = obj->un_pager.swp.swp_priv; 421 if (shm == NULL) 422 return; 423 if (!vm_pager_has_page(obj, m->pindex, NULL, NULL)) { 424 KASSERT(shm->shm_pages >= 1, 425 ("shm %p pages %jd free 1", shm, 426 (uintmax_t)shm->shm_pages)); 427 shm->shm_pages -= 1; 428 } 429 } 430 431 static struct pagerops shm_swap_pager_ops = { 432 .pgo_kvme_type = KVME_TYPE_SWAP, 433 .pgo_freespace = shm_pager_freespace, 434 .pgo_page_inserted = shm_page_inserted, 435 .pgo_page_removed = shm_page_removed, 436 }; 437 static int shmfd_pager_type = -1; 438 439 static int 440 shm_seek(struct file *fp, off_t offset, int whence, struct thread *td) 441 { 442 struct shmfd *shmfd; 443 off_t foffset; 444 int error; 445 446 shmfd = fp->f_data; 447 foffset = foffset_lock(fp, 0); 448 error = 0; 449 switch (whence) { 450 case L_INCR: 451 if (foffset < 0 || 452 (offset > 0 && foffset > OFF_MAX - offset)) { 453 error = EOVERFLOW; 454 break; 455 } 456 offset += foffset; 457 break; 458 case L_XTND: 459 if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) { 460 error = EOVERFLOW; 461 break; 462 } 463 offset += shmfd->shm_size; 464 break; 465 case L_SET: 466 break; 467 default: 468 error = EINVAL; 469 } 470 if (error == 0) { 471 if (offset < 0 || offset > shmfd->shm_size) 472 error = EINVAL; 473 else 474 td->td_uretoff.tdu_off = offset; 475 } 476 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0); 477 return (error); 478 } 479 480 static int 481 shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred, 482 int flags, struct thread *td) 483 { 484 struct shmfd *shmfd; 485 void *rl_cookie; 486 int error; 487 488 shmfd = fp->f_data; 489 #ifdef MAC 490 error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd); 491 if (error) 492 return (error); 493 #endif 494 foffset_lock_uio(fp, uio, flags); 495 rl_cookie = shm_rangelock_rlock(shmfd, uio->uio_offset, 496 uio->uio_offset + uio->uio_resid); 497 error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio); 498 shm_rangelock_unlock(shmfd, rl_cookie); 499 foffset_unlock_uio(fp, uio, flags); 500 return (error); 501 } 502 503 static int 504 shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred, 505 int flags, struct thread *td) 506 { 507 struct shmfd *shmfd; 508 void *rl_cookie; 509 int error; 510 off_t newsize; 511 512 KASSERT((flags & FOF_OFFSET) == 0 || uio->uio_offset >= 0, 513 ("%s: negative offset", __func__)); 514 515 shmfd = fp->f_data; 516 #ifdef MAC 517 error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd); 518 if (error) 519 return (error); 520 #endif 521 foffset_lock_uio(fp, uio, flags); 522 if (uio->uio_resid > OFF_MAX - uio->uio_offset) { 523 /* 524 * Overflow is only an error if we're supposed to expand on 525 * write. Otherwise, we'll just truncate the write to the 526 * size of the file, which can only grow up to OFF_MAX. 527 */ 528 if ((shmfd->shm_flags & SHM_GROW_ON_WRITE) != 0) { 529 foffset_unlock_uio(fp, uio, flags); 530 return (EFBIG); 531 } 532 533 newsize = atomic_load_64(&shmfd->shm_size); 534 } else { 535 newsize = uio->uio_offset + uio->uio_resid; 536 } 537 if ((flags & FOF_OFFSET) == 0) 538 rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX); 539 else 540 rl_cookie = shm_rangelock_wlock(shmfd, uio->uio_offset, 541 MAX(newsize, uio->uio_offset)); 542 if (shm_largepage(shmfd) && shmfd->shm_lp_psind == 0) { 543 error = EINVAL; 544 } else if ((shmfd->shm_seals & F_SEAL_WRITE) != 0) { 545 error = EPERM; 546 } else { 547 error = 0; 548 if ((shmfd->shm_flags & SHM_GROW_ON_WRITE) != 0 && 549 newsize > shmfd->shm_size) { 550 error = shm_dotruncate_cookie(shmfd, newsize, 551 rl_cookie); 552 } 553 if (error == 0) 554 error = uiomove_object(shmfd->shm_object, 555 shmfd->shm_size, uio); 556 } 557 shm_rangelock_unlock(shmfd, rl_cookie); 558 foffset_unlock_uio(fp, uio, flags); 559 return (error); 560 } 561 562 static int 563 shm_truncate(struct file *fp, off_t length, struct ucred *active_cred, 564 struct thread *td) 565 { 566 struct shmfd *shmfd; 567 #ifdef MAC 568 int error; 569 #endif 570 571 shmfd = fp->f_data; 572 #ifdef MAC 573 error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd); 574 if (error) 575 return (error); 576 #endif 577 return (shm_dotruncate(shmfd, length)); 578 } 579 580 int 581 shm_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred, 582 struct thread *td) 583 { 584 struct shmfd *shmfd; 585 struct shm_largepage_conf *conf; 586 void *rl_cookie; 587 588 shmfd = fp->f_data; 589 switch (com) { 590 case FIONBIO: 591 case FIOASYNC: 592 /* 593 * Allow fcntl(fd, F_SETFL, O_NONBLOCK) to work, 594 * just like it would on an unlinked regular file 595 */ 596 return (0); 597 case FIOSSHMLPGCNF: 598 if (!shm_largepage(shmfd)) 599 return (ENOTTY); 600 conf = data; 601 rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX); 602 if (shmfd->shm_lp_psind != 0 && 603 conf->psind != shmfd->shm_lp_psind) { 604 shm_rangelock_unlock(shmfd, rl_cookie); 605 return (EINVAL); 606 } 607 if (conf->psind <= 0 || conf->psind >= MAXPAGESIZES || 608 pagesizes[conf->psind] == 0) { 609 shm_rangelock_unlock(shmfd, rl_cookie); 610 return (EINVAL); 611 } 612 if (conf->alloc_policy != SHM_LARGEPAGE_ALLOC_DEFAULT && 613 conf->alloc_policy != SHM_LARGEPAGE_ALLOC_NOWAIT && 614 conf->alloc_policy != SHM_LARGEPAGE_ALLOC_HARD) { 615 shm_rangelock_unlock(shmfd, rl_cookie); 616 return (EINVAL); 617 } 618 shmfd->shm_lp_psind = conf->psind; 619 shmfd->shm_lp_alloc_policy = conf->alloc_policy; 620 shmfd->shm_object->un_pager.phys.data_val = conf->psind; 621 shm_rangelock_unlock(shmfd, rl_cookie); 622 return (0); 623 case FIOGSHMLPGCNF: 624 if (!shm_largepage(shmfd)) 625 return (ENOTTY); 626 conf = data; 627 rl_cookie = shm_rangelock_rlock(shmfd, 0, OFF_MAX); 628 conf->psind = shmfd->shm_lp_psind; 629 conf->alloc_policy = shmfd->shm_lp_alloc_policy; 630 shm_rangelock_unlock(shmfd, rl_cookie); 631 return (0); 632 default: 633 return (ENOTTY); 634 } 635 } 636 637 static int 638 shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred) 639 { 640 struct shmfd *shmfd; 641 #ifdef MAC 642 int error; 643 #endif 644 645 shmfd = fp->f_data; 646 647 #ifdef MAC 648 error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd); 649 if (error) 650 return (error); 651 #endif 652 653 /* 654 * Attempt to return sanish values for fstat() on a memory file 655 * descriptor. 656 */ 657 bzero(sb, sizeof(*sb)); 658 sb->st_blksize = PAGE_SIZE; 659 sb->st_size = shmfd->shm_size; 660 mtx_lock(&shm_timestamp_lock); 661 sb->st_atim = shmfd->shm_atime; 662 sb->st_ctim = shmfd->shm_ctime; 663 sb->st_mtim = shmfd->shm_mtime; 664 sb->st_birthtim = shmfd->shm_birthtime; 665 sb->st_mode = S_IFREG | shmfd->shm_mode; /* XXX */ 666 sb->st_uid = shmfd->shm_uid; 667 sb->st_gid = shmfd->shm_gid; 668 mtx_unlock(&shm_timestamp_lock); 669 sb->st_dev = shm_dev_ino; 670 sb->st_ino = shmfd->shm_ino; 671 sb->st_nlink = shmfd->shm_object->ref_count; 672 sb->st_blocks = ptoa(shm_largepage(shmfd) ? shmfd->shm_object->size : 673 shmfd->shm_pages) / DEV_BSIZE; 674 675 return (0); 676 } 677 678 static int 679 shm_close(struct file *fp, struct thread *td) 680 { 681 struct shmfd *shmfd; 682 683 shmfd = fp->f_data; 684 fp->f_data = NULL; 685 shm_drop(shmfd); 686 687 return (0); 688 } 689 690 static int 691 shm_copyin_path(struct thread *td, const char *userpath_in, char **path_out) { 692 int error; 693 char *path; 694 const char *pr_path; 695 size_t pr_pathlen; 696 697 path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK); 698 pr_path = td->td_ucred->cr_prison->pr_path; 699 700 /* Construct a full pathname for jailed callers. */ 701 pr_pathlen = strcmp(pr_path, "/") == 702 0 ? 0 : strlcpy(path, pr_path, MAXPATHLEN); 703 error = copyinstr(userpath_in, path + pr_pathlen, 704 MAXPATHLEN - pr_pathlen, NULL); 705 if (error != 0) 706 goto out; 707 708 #ifdef KTRACE 709 if (KTRPOINT(curthread, KTR_NAMEI)) 710 ktrnamei(path); 711 #endif 712 713 /* Require paths to start with a '/' character. */ 714 if (path[pr_pathlen] != '/') { 715 error = EINVAL; 716 goto out; 717 } 718 719 *path_out = path; 720 721 out: 722 if (error != 0) 723 free(path, M_SHMFD); 724 725 return (error); 726 } 727 728 static int 729 shm_partial_page_invalidate(vm_object_t object, vm_pindex_t idx, int base, 730 int end) 731 { 732 return (vm_page_grab_zero_partial(object, idx, base, end)); 733 } 734 735 static int 736 shm_dotruncate_locked(struct shmfd *shmfd, off_t length, void *rl_cookie) 737 { 738 vm_object_t object; 739 vm_pindex_t nobjsize; 740 vm_ooffset_t delta; 741 int base, error; 742 743 KASSERT(length >= 0, ("shm_dotruncate: length < 0")); 744 object = shmfd->shm_object; 745 VM_OBJECT_ASSERT_WLOCKED(object); 746 rangelock_cookie_assert(rl_cookie, RCA_WLOCKED); 747 if (length == shmfd->shm_size) 748 return (0); 749 nobjsize = OFF_TO_IDX(length + PAGE_MASK); 750 751 /* Are we shrinking? If so, trim the end. */ 752 if (length < shmfd->shm_size) { 753 if ((shmfd->shm_seals & F_SEAL_SHRINK) != 0) 754 return (EPERM); 755 756 /* 757 * Disallow any requests to shrink the size if this 758 * object is mapped into the kernel. 759 */ 760 if (shmfd->shm_kmappings > 0) 761 return (EBUSY); 762 763 /* 764 * Zero the truncated part of the last page. 765 */ 766 base = length & PAGE_MASK; 767 if (base != 0) { 768 error = shm_partial_page_invalidate(object, 769 OFF_TO_IDX(length), base, PAGE_SIZE); 770 if (error) 771 return (error); 772 } 773 delta = IDX_TO_OFF(object->size - nobjsize); 774 775 if (nobjsize < object->size) 776 vm_object_page_remove(object, nobjsize, object->size, 777 0); 778 779 /* Free the swap accounted for shm */ 780 swap_release_by_cred(delta, object->cred); 781 } else { 782 if ((shmfd->shm_seals & F_SEAL_GROW) != 0) 783 return (EPERM); 784 785 /* Try to reserve additional swap space. */ 786 delta = IDX_TO_OFF(nobjsize - object->size); 787 if (!swap_reserve_by_cred(delta, object->cred)) 788 return (ENOMEM); 789 } 790 shmfd->shm_size = length; 791 mtx_lock(&shm_timestamp_lock); 792 vfs_timestamp(&shmfd->shm_ctime); 793 shmfd->shm_mtime = shmfd->shm_ctime; 794 mtx_unlock(&shm_timestamp_lock); 795 object->size = nobjsize; 796 return (0); 797 } 798 799 static int 800 shm_dotruncate_largepage(struct shmfd *shmfd, off_t length, void *rl_cookie) 801 { 802 vm_object_t object; 803 vm_page_t m; 804 vm_pindex_t newobjsz; 805 vm_pindex_t oldobjsz __unused; 806 int aflags, error, i, psind, try; 807 808 KASSERT(length >= 0, ("shm_dotruncate_largepage: length < 0")); 809 object = shmfd->shm_object; 810 VM_OBJECT_ASSERT_WLOCKED(object); 811 rangelock_cookie_assert(rl_cookie, RCA_WLOCKED); 812 813 oldobjsz = object->size; 814 newobjsz = OFF_TO_IDX(length); 815 if (length == shmfd->shm_size) 816 return (0); 817 psind = shmfd->shm_lp_psind; 818 if (psind == 0 && length != 0) 819 return (EINVAL); 820 if ((length & (pagesizes[psind] - 1)) != 0) 821 return (EINVAL); 822 823 if (length < shmfd->shm_size) { 824 if ((shmfd->shm_seals & F_SEAL_SHRINK) != 0) 825 return (EPERM); 826 if (shmfd->shm_kmappings > 0) 827 return (EBUSY); 828 return (ENOTSUP); /* Pages are unmanaged. */ 829 #if 0 830 vm_object_page_remove(object, newobjsz, oldobjsz, 0); 831 object->size = newobjsz; 832 shmfd->shm_size = length; 833 return (0); 834 #endif 835 } 836 837 if ((shmfd->shm_seals & F_SEAL_GROW) != 0) 838 return (EPERM); 839 840 aflags = VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_WIRED; 841 if (shmfd->shm_lp_alloc_policy == SHM_LARGEPAGE_ALLOC_NOWAIT) 842 aflags |= VM_ALLOC_WAITFAIL; 843 try = 0; 844 845 /* 846 * Extend shmfd and object, keeping all already fully 847 * allocated large pages intact even on error, because dropped 848 * object lock might allowed mapping of them. 849 */ 850 while (object->size < newobjsz) { 851 error = sig_intr(); 852 if (error != 0) 853 return (error); 854 m = vm_page_alloc_contig(object, object->size, aflags, 855 pagesizes[psind] / PAGE_SIZE, 0, ~0, 856 pagesizes[psind], 0, 857 VM_MEMATTR_DEFAULT); 858 if (m == NULL) { 859 VM_OBJECT_WUNLOCK(object); 860 error = sig_intr(); 861 if (error != 0) { 862 VM_OBJECT_WLOCK(object); 863 return (error); 864 } 865 if (shmfd->shm_lp_alloc_policy == 866 SHM_LARGEPAGE_ALLOC_NOWAIT || 867 (shmfd->shm_lp_alloc_policy == 868 SHM_LARGEPAGE_ALLOC_DEFAULT && 869 try >= largepage_reclaim_tries)) { 870 VM_OBJECT_WLOCK(object); 871 return (ENOMEM); 872 } 873 error = vm_page_reclaim_contig(aflags, 874 pagesizes[psind] / PAGE_SIZE, 0, ~0, 875 pagesizes[psind], 0); 876 if (error == ENOMEM) 877 error = vm_wait_intr(object); 878 if (error != 0) { 879 VM_OBJECT_WLOCK(object); 880 return (error); 881 } 882 try++; 883 VM_OBJECT_WLOCK(object); 884 continue; 885 } 886 try = 0; 887 for (i = 0; i < pagesizes[psind] / PAGE_SIZE; i++) { 888 if ((m[i].flags & PG_ZERO) == 0) 889 pmap_zero_page(&m[i]); 890 vm_page_valid(&m[i]); 891 vm_page_xunbusy(&m[i]); 892 } 893 object->size += OFF_TO_IDX(pagesizes[psind]); 894 shmfd->shm_size += pagesizes[psind]; 895 atomic_add_long(&count_largepages[psind], 1); 896 } 897 return (0); 898 } 899 900 static int 901 shm_dotruncate_cookie(struct shmfd *shmfd, off_t length, void *rl_cookie) 902 { 903 int error; 904 905 VM_OBJECT_WLOCK(shmfd->shm_object); 906 error = shm_largepage(shmfd) ? shm_dotruncate_largepage(shmfd, 907 length, rl_cookie) : shm_dotruncate_locked(shmfd, length, 908 rl_cookie); 909 VM_OBJECT_WUNLOCK(shmfd->shm_object); 910 return (error); 911 } 912 913 int 914 shm_dotruncate(struct shmfd *shmfd, off_t length) 915 { 916 void *rl_cookie; 917 int error; 918 919 rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX); 920 error = shm_dotruncate_cookie(shmfd, length, rl_cookie); 921 shm_rangelock_unlock(shmfd, rl_cookie); 922 return (error); 923 } 924 925 /* 926 * shmfd object management including creation and reference counting 927 * routines. 928 */ 929 struct shmfd * 930 shm_alloc(struct ucred *ucred, mode_t mode, bool largepage) 931 { 932 struct shmfd *shmfd; 933 vm_object_t obj; 934 935 if (largepage) { 936 obj = phys_pager_allocate(NULL, &shm_largepage_phys_ops, 937 NULL, 0, VM_PROT_DEFAULT, 0, ucred); 938 } else { 939 obj = vm_pager_allocate(shmfd_pager_type, NULL, 0, 940 VM_PROT_DEFAULT, 0, ucred); 941 } 942 if (obj == NULL) { 943 /* 944 * swap reservation limits can cause object allocation 945 * to fail. 946 */ 947 return (NULL); 948 } 949 950 shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO); 951 shmfd->shm_uid = ucred->cr_uid; 952 shmfd->shm_gid = ucred->cr_gid; 953 shmfd->shm_mode = mode; 954 if (largepage) { 955 obj->un_pager.phys.phys_priv = shmfd; 956 shmfd->shm_lp_alloc_policy = SHM_LARGEPAGE_ALLOC_DEFAULT; 957 } else { 958 obj->un_pager.swp.swp_priv = shmfd; 959 } 960 961 VM_OBJECT_WLOCK(obj); 962 vm_object_set_flag(obj, OBJ_POSIXSHM); 963 VM_OBJECT_WUNLOCK(obj); 964 shmfd->shm_object = obj; 965 vfs_timestamp(&shmfd->shm_birthtime); 966 shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime = 967 shmfd->shm_birthtime; 968 shmfd->shm_ino = alloc_unr64(&shm_ino_unr); 969 refcount_init(&shmfd->shm_refs, 1); 970 mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF); 971 rangelock_init(&shmfd->shm_rl); 972 #ifdef MAC 973 mac_posixshm_init(shmfd); 974 mac_posixshm_create(ucred, shmfd); 975 #endif 976 977 return (shmfd); 978 } 979 980 struct shmfd * 981 shm_hold(struct shmfd *shmfd) 982 { 983 984 refcount_acquire(&shmfd->shm_refs); 985 return (shmfd); 986 } 987 988 void 989 shm_drop(struct shmfd *shmfd) 990 { 991 vm_object_t obj; 992 993 if (refcount_release(&shmfd->shm_refs)) { 994 #ifdef MAC 995 mac_posixshm_destroy(shmfd); 996 #endif 997 rangelock_destroy(&shmfd->shm_rl); 998 mtx_destroy(&shmfd->shm_mtx); 999 obj = shmfd->shm_object; 1000 VM_OBJECT_WLOCK(obj); 1001 if (shm_largepage(shmfd)) 1002 obj->un_pager.phys.phys_priv = NULL; 1003 else 1004 obj->un_pager.swp.swp_priv = NULL; 1005 VM_OBJECT_WUNLOCK(obj); 1006 vm_object_deallocate(obj); 1007 free(shmfd, M_SHMFD); 1008 } 1009 } 1010 1011 /* 1012 * Determine if the credentials have sufficient permissions for a 1013 * specified combination of FREAD and FWRITE. 1014 */ 1015 int 1016 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags) 1017 { 1018 accmode_t accmode; 1019 int error; 1020 1021 accmode = 0; 1022 if (flags & FREAD) 1023 accmode |= VREAD; 1024 if (flags & FWRITE) 1025 accmode |= VWRITE; 1026 mtx_lock(&shm_timestamp_lock); 1027 error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid, 1028 accmode, ucred); 1029 mtx_unlock(&shm_timestamp_lock); 1030 return (error); 1031 } 1032 1033 static void 1034 shm_init(void *arg) 1035 { 1036 char name[32]; 1037 int i; 1038 1039 mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF); 1040 sx_init(&shm_dict_lock, "shm dictionary"); 1041 shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash); 1042 new_unrhdr64(&shm_ino_unr, 1); 1043 shm_dev_ino = devfs_alloc_cdp_inode(); 1044 KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized")); 1045 shmfd_pager_type = vm_pager_alloc_dyn_type(&shm_swap_pager_ops, 1046 OBJT_SWAP); 1047 MPASS(shmfd_pager_type != -1); 1048 1049 for (i = 1; i < MAXPAGESIZES; i++) { 1050 if (pagesizes[i] == 0) 1051 break; 1052 #define M (1024 * 1024) 1053 #define G (1024 * M) 1054 if (pagesizes[i] >= G) 1055 snprintf(name, sizeof(name), "%luG", pagesizes[i] / G); 1056 else if (pagesizes[i] >= M) 1057 snprintf(name, sizeof(name), "%luM", pagesizes[i] / M); 1058 else 1059 snprintf(name, sizeof(name), "%lu", pagesizes[i]); 1060 #undef G 1061 #undef M 1062 SYSCTL_ADD_ULONG(NULL, SYSCTL_STATIC_CHILDREN(_vm_largepages), 1063 OID_AUTO, name, CTLFLAG_RD, &count_largepages[i], 1064 "number of non-transient largepages allocated"); 1065 } 1066 } 1067 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL); 1068 1069 /* 1070 * Remove all shared memory objects that belong to a prison. 1071 */ 1072 void 1073 shm_remove_prison(struct prison *pr) 1074 { 1075 struct shm_mapping *shmm, *tshmm; 1076 u_long i; 1077 1078 sx_xlock(&shm_dict_lock); 1079 for (i = 0; i < shm_hash + 1; i++) { 1080 LIST_FOREACH_SAFE(shmm, &shm_dictionary[i], sm_link, tshmm) { 1081 if (shmm->sm_shmfd->shm_object->cred && 1082 shmm->sm_shmfd->shm_object->cred->cr_prison == pr) 1083 shm_doremove(shmm); 1084 } 1085 } 1086 sx_xunlock(&shm_dict_lock); 1087 } 1088 1089 /* 1090 * Dictionary management. We maintain an in-kernel dictionary to map 1091 * paths to shmfd objects. We use the FNV hash on the path to store 1092 * the mappings in a hash table. 1093 */ 1094 static struct shmfd * 1095 shm_lookup(char *path, Fnv32_t fnv) 1096 { 1097 struct shm_mapping *map; 1098 1099 LIST_FOREACH(map, SHM_HASH(fnv), sm_link) { 1100 if (map->sm_fnv != fnv) 1101 continue; 1102 if (strcmp(map->sm_path, path) == 0) 1103 return (map->sm_shmfd); 1104 } 1105 1106 return (NULL); 1107 } 1108 1109 static void 1110 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd) 1111 { 1112 struct shm_mapping *map; 1113 1114 map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK); 1115 map->sm_path = path; 1116 map->sm_fnv = fnv; 1117 map->sm_shmfd = shm_hold(shmfd); 1118 shmfd->shm_path = path; 1119 LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link); 1120 } 1121 1122 static int 1123 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred) 1124 { 1125 struct shm_mapping *map; 1126 int error; 1127 1128 LIST_FOREACH(map, SHM_HASH(fnv), sm_link) { 1129 if (map->sm_fnv != fnv) 1130 continue; 1131 if (strcmp(map->sm_path, path) == 0) { 1132 #ifdef MAC 1133 error = mac_posixshm_check_unlink(ucred, map->sm_shmfd); 1134 if (error) 1135 return (error); 1136 #endif 1137 error = shm_access(map->sm_shmfd, ucred, 1138 FREAD | FWRITE); 1139 if (error) 1140 return (error); 1141 shm_doremove(map); 1142 return (0); 1143 } 1144 } 1145 1146 return (ENOENT); 1147 } 1148 1149 static void 1150 shm_doremove(struct shm_mapping *map) 1151 { 1152 map->sm_shmfd->shm_path = NULL; 1153 LIST_REMOVE(map, sm_link); 1154 shm_drop(map->sm_shmfd); 1155 free(map->sm_path, M_SHMFD); 1156 free(map, M_SHMFD); 1157 } 1158 1159 int 1160 kern_shm_open2(struct thread *td, const char *userpath, int flags, mode_t mode, 1161 int shmflags, struct filecaps *fcaps, const char *name __unused, 1162 struct shmfd *shmfd) 1163 { 1164 struct pwddesc *pdp; 1165 struct file *fp; 1166 char *path; 1167 void *rl_cookie; 1168 Fnv32_t fnv; 1169 mode_t cmode; 1170 int error, fd, initial_seals; 1171 bool largepage; 1172 1173 if ((shmflags & ~(SHM_ALLOW_SEALING | SHM_GROW_ON_WRITE | 1174 SHM_LARGEPAGE)) != 0) 1175 return (EINVAL); 1176 1177 initial_seals = F_SEAL_SEAL; 1178 if ((shmflags & SHM_ALLOW_SEALING) != 0) 1179 initial_seals &= ~F_SEAL_SEAL; 1180 1181 AUDIT_ARG_FFLAGS(flags); 1182 AUDIT_ARG_MODE(mode); 1183 1184 if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR) 1185 return (EINVAL); 1186 1187 if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC | 1188 O_CLOFORK)) != 0) 1189 return (EINVAL); 1190 1191 largepage = (shmflags & SHM_LARGEPAGE) != 0; 1192 if (largepage && !PMAP_HAS_LARGEPAGES) 1193 return (ENOTTY); 1194 1195 /* 1196 * Currently only F_SEAL_SEAL may be set when creating or opening shmfd. 1197 * If the decision is made later to allow additional seals, care must be 1198 * taken below to ensure that the seals are properly set if the shmfd 1199 * already existed -- this currently assumes that only F_SEAL_SEAL can 1200 * be set and doesn't take further precautions to ensure the validity of 1201 * the seals being added with respect to current mappings. 1202 */ 1203 if ((initial_seals & ~F_SEAL_SEAL) != 0) 1204 return (EINVAL); 1205 1206 if (userpath != SHM_ANON) { 1207 error = shm_copyin_path(td, userpath, &path); 1208 if (error != 0) 1209 return (error); 1210 1211 #ifdef CAPABILITY_MODE 1212 /* 1213 * shm_open(2) is only allowed for anonymous objects. 1214 */ 1215 if (CAP_TRACING(td)) 1216 ktrcapfail(CAPFAIL_NAMEI, path); 1217 if (IN_CAPABILITY_MODE(td)) { 1218 error = ECAPMODE; 1219 goto outnofp; 1220 } 1221 #endif 1222 1223 AUDIT_ARG_UPATH1_CANON(path); 1224 } else { 1225 path = NULL; 1226 } 1227 1228 pdp = td->td_proc->p_pd; 1229 cmode = (mode & ~pdp->pd_cmask) & ACCESSPERMS; 1230 1231 /* 1232 * shm_open(2) created shm should always have O_CLOEXEC set, as mandated 1233 * by POSIX. We allow it to be unset here so that an in-kernel 1234 * interface may be written as a thin layer around shm, optionally not 1235 * setting CLOEXEC. For shm_open(2), O_CLOEXEC is set unconditionally 1236 * in sys_shm_open() to keep this implementation compliant. 1237 */ 1238 error = falloc_caps(td, &fp, &fd, flags & O_CLOEXEC, fcaps); 1239 if (error != 0) 1240 goto outnofp; 1241 1242 /* 1243 * A SHM_ANON path pointer creates an anonymous object. We allow other 1244 * parts of the kernel to pre-populate a shmfd and then materialize an 1245 * fd for it here as a means to pass data back up to userland. This 1246 * doesn't really make sense for named shm objects, but it makes plenty 1247 * of sense for anonymous objects. 1248 */ 1249 if (userpath == SHM_ANON) { 1250 if (shmfd != NULL) { 1251 shm_hold(shmfd); 1252 } else { 1253 /* 1254 * A read-only anonymous object is pointless, unless it 1255 * was pre-populated by the kernel with the expectation 1256 * that a shmfd would later be created for userland to 1257 * access it through. 1258 */ 1259 if ((flags & O_ACCMODE) == O_RDONLY) { 1260 error = EINVAL; 1261 goto out; 1262 } 1263 shmfd = shm_alloc(td->td_ucred, cmode, largepage); 1264 if (shmfd == NULL) { 1265 error = ENOMEM; 1266 goto out; 1267 } 1268 1269 shmfd->shm_seals = initial_seals; 1270 shmfd->shm_flags = shmflags; 1271 } 1272 } else { 1273 fnv = fnv_32_str(path, FNV1_32_INIT); 1274 sx_xlock(&shm_dict_lock); 1275 1276 MPASS(shmfd == NULL); 1277 shmfd = shm_lookup(path, fnv); 1278 if (shmfd == NULL) { 1279 /* Object does not yet exist, create it if requested. */ 1280 if (flags & O_CREAT) { 1281 #ifdef MAC 1282 error = mac_posixshm_check_create(td->td_ucred, 1283 path); 1284 if (error == 0) { 1285 #endif 1286 shmfd = shm_alloc(td->td_ucred, cmode, 1287 largepage); 1288 if (shmfd == NULL) { 1289 error = ENOMEM; 1290 } else { 1291 shmfd->shm_seals = 1292 initial_seals; 1293 shmfd->shm_flags = shmflags; 1294 shm_insert(path, fnv, shmfd); 1295 path = NULL; 1296 } 1297 #ifdef MAC 1298 } 1299 #endif 1300 } else { 1301 error = ENOENT; 1302 } 1303 } else { 1304 /* 1305 * Object already exists, obtain a new reference if 1306 * requested and permitted. 1307 */ 1308 rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX); 1309 1310 /* 1311 * kern_shm_open() likely shouldn't ever error out on 1312 * trying to set a seal that already exists, unlike 1313 * F_ADD_SEALS. This would break terribly as 1314 * shm_open(2) actually sets F_SEAL_SEAL to maintain 1315 * historical behavior where the underlying file could 1316 * not be sealed. 1317 */ 1318 initial_seals &= ~shmfd->shm_seals; 1319 1320 /* 1321 * initial_seals can't set additional seals if we've 1322 * already been set F_SEAL_SEAL. If F_SEAL_SEAL is set, 1323 * then we've already removed that one from 1324 * initial_seals. This is currently redundant as we 1325 * only allow setting F_SEAL_SEAL at creation time, but 1326 * it's cheap to check and decreases the effort required 1327 * to allow additional seals. 1328 */ 1329 if ((shmfd->shm_seals & F_SEAL_SEAL) != 0 && 1330 initial_seals != 0) 1331 error = EPERM; 1332 else if ((flags & (O_CREAT | O_EXCL)) == 1333 (O_CREAT | O_EXCL)) 1334 error = EEXIST; 1335 else if (shmflags != 0 && shmflags != shmfd->shm_flags) 1336 error = EINVAL; 1337 else { 1338 #ifdef MAC 1339 error = mac_posixshm_check_open(td->td_ucred, 1340 shmfd, FFLAGS(flags & O_ACCMODE)); 1341 if (error == 0) 1342 #endif 1343 error = shm_access(shmfd, td->td_ucred, 1344 FFLAGS(flags & O_ACCMODE)); 1345 } 1346 1347 /* 1348 * Truncate the file back to zero length if 1349 * O_TRUNC was specified and the object was 1350 * opened with read/write. 1351 */ 1352 if (error == 0 && 1353 (flags & (O_ACCMODE | O_TRUNC)) == 1354 (O_RDWR | O_TRUNC)) { 1355 #ifdef MAC 1356 error = mac_posixshm_check_truncate( 1357 td->td_ucred, fp->f_cred, shmfd); 1358 if (error == 0) 1359 #endif 1360 error = shm_dotruncate_cookie(shmfd, 0, 1361 rl_cookie); 1362 } 1363 if (error == 0) { 1364 /* 1365 * Currently we only allow F_SEAL_SEAL to be 1366 * set initially. As noted above, this would 1367 * need to be reworked should that change. 1368 */ 1369 shmfd->shm_seals |= initial_seals; 1370 shm_hold(shmfd); 1371 } 1372 shm_rangelock_unlock(shmfd, rl_cookie); 1373 } 1374 sx_xunlock(&shm_dict_lock); 1375 1376 if (error != 0) 1377 goto out; 1378 } 1379 1380 finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops); 1381 1382 td->td_retval[0] = fd; 1383 fdrop(fp, td); 1384 free(path, M_SHMFD); 1385 1386 return (0); 1387 1388 out: 1389 fdclose(td, fp, fd); 1390 fdrop(fp, td); 1391 outnofp: 1392 free(path, M_SHMFD); 1393 1394 return (error); 1395 } 1396 1397 /* System calls. */ 1398 #ifdef COMPAT_FREEBSD12 1399 int 1400 freebsd12_shm_open(struct thread *td, struct freebsd12_shm_open_args *uap) 1401 { 1402 1403 return (kern_shm_open(td, uap->path, uap->flags | O_CLOEXEC, 1404 uap->mode, NULL)); 1405 } 1406 #endif 1407 1408 int 1409 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap) 1410 { 1411 char *path; 1412 Fnv32_t fnv; 1413 int error; 1414 1415 error = shm_copyin_path(td, uap->path, &path); 1416 if (error != 0) 1417 return (error); 1418 1419 AUDIT_ARG_UPATH1_CANON(path); 1420 fnv = fnv_32_str(path, FNV1_32_INIT); 1421 sx_xlock(&shm_dict_lock); 1422 error = shm_remove(path, fnv, td->td_ucred); 1423 sx_xunlock(&shm_dict_lock); 1424 free(path, M_SHMFD); 1425 1426 return (error); 1427 } 1428 1429 int 1430 sys_shm_rename(struct thread *td, struct shm_rename_args *uap) 1431 { 1432 char *path_from = NULL, *path_to = NULL; 1433 Fnv32_t fnv_from, fnv_to; 1434 struct shmfd *fd_from; 1435 struct shmfd *fd_to; 1436 int error; 1437 int flags; 1438 1439 flags = uap->flags; 1440 AUDIT_ARG_FFLAGS(flags); 1441 1442 /* 1443 * Make sure the user passed only valid flags. 1444 * If you add a new flag, please add a new term here. 1445 */ 1446 if ((flags & ~( 1447 SHM_RENAME_NOREPLACE | 1448 SHM_RENAME_EXCHANGE 1449 )) != 0) { 1450 error = EINVAL; 1451 goto out; 1452 } 1453 1454 /* 1455 * EXCHANGE and NOREPLACE don't quite make sense together. Let's 1456 * force the user to choose one or the other. 1457 */ 1458 if ((flags & SHM_RENAME_NOREPLACE) != 0 && 1459 (flags & SHM_RENAME_EXCHANGE) != 0) { 1460 error = EINVAL; 1461 goto out; 1462 } 1463 1464 /* Renaming to or from anonymous makes no sense */ 1465 if (uap->path_from == SHM_ANON || uap->path_to == SHM_ANON) { 1466 error = EINVAL; 1467 goto out; 1468 } 1469 1470 error = shm_copyin_path(td, uap->path_from, &path_from); 1471 if (error != 0) 1472 goto out; 1473 1474 error = shm_copyin_path(td, uap->path_to, &path_to); 1475 if (error != 0) 1476 goto out; 1477 1478 AUDIT_ARG_UPATH1_CANON(path_from); 1479 AUDIT_ARG_UPATH2_CANON(path_to); 1480 1481 /* Rename with from/to equal is a no-op */ 1482 if (strcmp(path_from, path_to) == 0) 1483 goto out; 1484 1485 fnv_from = fnv_32_str(path_from, FNV1_32_INIT); 1486 fnv_to = fnv_32_str(path_to, FNV1_32_INIT); 1487 1488 sx_xlock(&shm_dict_lock); 1489 1490 fd_from = shm_lookup(path_from, fnv_from); 1491 if (fd_from == NULL) { 1492 error = ENOENT; 1493 goto out_locked; 1494 } 1495 1496 fd_to = shm_lookup(path_to, fnv_to); 1497 if ((flags & SHM_RENAME_NOREPLACE) != 0 && fd_to != NULL) { 1498 error = EEXIST; 1499 goto out_locked; 1500 } 1501 1502 /* 1503 * Unconditionally prevents shm_remove from invalidating the 'from' 1504 * shm's state. 1505 */ 1506 shm_hold(fd_from); 1507 error = shm_remove(path_from, fnv_from, td->td_ucred); 1508 1509 /* 1510 * One of my assumptions failed if ENOENT (e.g. locking didn't 1511 * protect us) 1512 */ 1513 KASSERT(error != ENOENT, ("Our shm disappeared during shm_rename: %s", 1514 path_from)); 1515 if (error != 0) { 1516 shm_drop(fd_from); 1517 goto out_locked; 1518 } 1519 1520 /* 1521 * If we are exchanging, we need to ensure the shm_remove below 1522 * doesn't invalidate the dest shm's state. 1523 */ 1524 if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) 1525 shm_hold(fd_to); 1526 1527 /* 1528 * NOTE: if path_to is not already in the hash, c'est la vie; 1529 * it simply means we have nothing already at path_to to unlink. 1530 * That is the ENOENT case. 1531 * 1532 * If we somehow don't have access to unlink this guy, but 1533 * did for the shm at path_from, then relink the shm to path_from 1534 * and abort with EACCES. 1535 * 1536 * All other errors: that is weird; let's relink and abort the 1537 * operation. 1538 */ 1539 error = shm_remove(path_to, fnv_to, td->td_ucred); 1540 if (error != 0 && error != ENOENT) { 1541 shm_insert(path_from, fnv_from, fd_from); 1542 shm_drop(fd_from); 1543 /* Don't free path_from now, since the hash references it */ 1544 path_from = NULL; 1545 goto out_locked; 1546 } 1547 1548 error = 0; 1549 1550 shm_insert(path_to, fnv_to, fd_from); 1551 1552 /* Don't free path_to now, since the hash references it */ 1553 path_to = NULL; 1554 1555 /* We kept a ref when we removed, and incremented again in insert */ 1556 shm_drop(fd_from); 1557 KASSERT(fd_from->shm_refs > 0, ("Expected >0 refs; got: %d\n", 1558 fd_from->shm_refs)); 1559 1560 if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) { 1561 shm_insert(path_from, fnv_from, fd_to); 1562 path_from = NULL; 1563 shm_drop(fd_to); 1564 KASSERT(fd_to->shm_refs > 0, ("Expected >0 refs; got: %d\n", 1565 fd_to->shm_refs)); 1566 } 1567 1568 out_locked: 1569 sx_xunlock(&shm_dict_lock); 1570 1571 out: 1572 free(path_from, M_SHMFD); 1573 free(path_to, M_SHMFD); 1574 return (error); 1575 } 1576 1577 static int 1578 shm_mmap_large(struct shmfd *shmfd, vm_map_t map, vm_offset_t *addr, 1579 vm_size_t size, vm_prot_t prot, vm_prot_t max_prot, int flags, 1580 vm_ooffset_t foff, struct thread *td, void *rl_cookie) 1581 { 1582 struct vmspace *vms; 1583 vm_map_entry_t next_entry, prev_entry; 1584 vm_offset_t align, mask, maxaddr; 1585 int docow, error, rv, try; 1586 bool curmap; 1587 1588 rangelock_cookie_assert(rl_cookie, RCA_LOCKED); 1589 1590 if (shmfd->shm_lp_psind == 0) 1591 return (EINVAL); 1592 1593 /* MAP_PRIVATE is disabled */ 1594 if ((flags & ~(MAP_SHARED | MAP_FIXED | MAP_EXCL | 1595 MAP_NOCORE | MAP_32BIT | MAP_ALIGNMENT_MASK)) != 0) 1596 return (EINVAL); 1597 1598 vms = td->td_proc->p_vmspace; 1599 curmap = map == &vms->vm_map; 1600 if (curmap) { 1601 error = kern_mmap_racct_check(td, map, size); 1602 if (error != 0) 1603 return (error); 1604 } 1605 1606 docow = shmfd->shm_lp_psind << MAP_SPLIT_BOUNDARY_SHIFT; 1607 docow |= MAP_INHERIT_SHARE; 1608 if ((flags & MAP_NOCORE) != 0) 1609 docow |= MAP_DISABLE_COREDUMP; 1610 1611 mask = pagesizes[shmfd->shm_lp_psind] - 1; 1612 if ((foff & mask) != 0) 1613 return (EINVAL); 1614 maxaddr = vm_map_max(map); 1615 if ((flags & MAP_32BIT) != 0 && maxaddr > MAP_32BIT_MAX_ADDR) 1616 maxaddr = MAP_32BIT_MAX_ADDR; 1617 if (size == 0 || (size & mask) != 0 || 1618 (*addr != 0 && ((*addr & mask) != 0 || 1619 *addr + size < *addr || *addr + size > maxaddr))) 1620 return (EINVAL); 1621 1622 align = flags & MAP_ALIGNMENT_MASK; 1623 if (align == 0) { 1624 align = pagesizes[shmfd->shm_lp_psind]; 1625 } else if (align == MAP_ALIGNED_SUPER) { 1626 /* 1627 * MAP_ALIGNED_SUPER is only supported on superpage sizes, 1628 * i.e., [1, VM_NRESERVLEVEL]. shmfd->shm_lp_psind < 1 is 1629 * handled above. 1630 */ 1631 if ( 1632 #if VM_NRESERVLEVEL > 0 1633 shmfd->shm_lp_psind > VM_NRESERVLEVEL 1634 #else 1635 shmfd->shm_lp_psind > 1 1636 #endif 1637 ) 1638 return (EINVAL); 1639 align = pagesizes[shmfd->shm_lp_psind]; 1640 } else { 1641 align >>= MAP_ALIGNMENT_SHIFT; 1642 align = 1ULL << align; 1643 /* Also handles overflow. */ 1644 if (align < pagesizes[shmfd->shm_lp_psind]) 1645 return (EINVAL); 1646 } 1647 1648 vm_map_lock(map); 1649 if ((flags & MAP_FIXED) == 0) { 1650 try = 1; 1651 if (curmap && (*addr == 0 || 1652 (*addr >= round_page((vm_offset_t)vms->vm_taddr) && 1653 *addr < round_page((vm_offset_t)vms->vm_daddr + 1654 lim_max(td, RLIMIT_DATA))))) { 1655 *addr = roundup2((vm_offset_t)vms->vm_daddr + 1656 lim_max(td, RLIMIT_DATA), 1657 pagesizes[shmfd->shm_lp_psind]); 1658 } 1659 again: 1660 rv = vm_map_find_aligned(map, addr, size, maxaddr, align); 1661 if (rv != KERN_SUCCESS) { 1662 if (try == 1) { 1663 try = 2; 1664 *addr = vm_map_min(map); 1665 if ((*addr & mask) != 0) 1666 *addr = (*addr + mask) & mask; 1667 goto again; 1668 } 1669 goto fail1; 1670 } 1671 } else if ((flags & MAP_EXCL) == 0) { 1672 rv = vm_map_delete(map, *addr, *addr + size); 1673 if (rv != KERN_SUCCESS) 1674 goto fail1; 1675 } else { 1676 error = ENOSPC; 1677 if (vm_map_lookup_entry(map, *addr, &prev_entry)) 1678 goto fail; 1679 next_entry = vm_map_entry_succ(prev_entry); 1680 if (next_entry->start < *addr + size) 1681 goto fail; 1682 } 1683 1684 rv = vm_map_insert(map, shmfd->shm_object, foff, *addr, *addr + size, 1685 prot, max_prot, docow); 1686 fail1: 1687 error = vm_mmap_to_errno(rv); 1688 fail: 1689 vm_map_unlock(map); 1690 return (error); 1691 } 1692 1693 static int 1694 shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize, 1695 vm_prot_t prot, vm_prot_t max_maxprot, int flags, 1696 vm_ooffset_t foff, struct thread *td) 1697 { 1698 struct shmfd *shmfd; 1699 vm_prot_t maxprot; 1700 int error; 1701 bool writecnt; 1702 void *rl_cookie; 1703 1704 shmfd = fp->f_data; 1705 maxprot = VM_PROT_NONE; 1706 1707 rl_cookie = shm_rangelock_rlock(shmfd, 0, objsize); 1708 /* FREAD should always be set. */ 1709 if ((fp->f_flag & FREAD) != 0) 1710 maxprot |= VM_PROT_EXECUTE | VM_PROT_READ; 1711 1712 /* 1713 * If FWRITE's set, we can allow VM_PROT_WRITE unless it's a shared 1714 * mapping with a write seal applied. Private mappings are always 1715 * writeable. 1716 */ 1717 if ((flags & MAP_SHARED) == 0) { 1718 if ((max_maxprot & VM_PROT_WRITE) != 0) 1719 maxprot |= VM_PROT_WRITE; 1720 writecnt = false; 1721 } else { 1722 if ((fp->f_flag & FWRITE) != 0 && 1723 (shmfd->shm_seals & F_SEAL_WRITE) == 0) 1724 maxprot |= VM_PROT_WRITE; 1725 1726 /* 1727 * Any mappings from a writable descriptor may be upgraded to 1728 * VM_PROT_WRITE with mprotect(2), unless a write-seal was 1729 * applied between the open and subsequent mmap(2). We want to 1730 * reject application of a write seal as long as any such 1731 * mapping exists so that the seal cannot be trivially bypassed. 1732 */ 1733 writecnt = (maxprot & VM_PROT_WRITE) != 0; 1734 if (!writecnt && (prot & VM_PROT_WRITE) != 0) { 1735 error = EACCES; 1736 goto out; 1737 } 1738 } 1739 maxprot &= max_maxprot; 1740 1741 /* See comment in vn_mmap(). */ 1742 if ( 1743 #ifdef _LP64 1744 objsize > OFF_MAX || 1745 #endif 1746 foff > OFF_MAX - objsize) { 1747 error = EINVAL; 1748 goto out; 1749 } 1750 1751 #ifdef MAC 1752 error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags); 1753 if (error != 0) 1754 goto out; 1755 #endif 1756 1757 mtx_lock(&shm_timestamp_lock); 1758 vfs_timestamp(&shmfd->shm_atime); 1759 mtx_unlock(&shm_timestamp_lock); 1760 vm_object_reference(shmfd->shm_object); 1761 1762 if (shm_largepage(shmfd)) { 1763 writecnt = false; 1764 error = shm_mmap_large(shmfd, map, addr, objsize, prot, 1765 maxprot, flags, foff, td, rl_cookie); 1766 } else { 1767 if (writecnt) { 1768 vm_pager_update_writecount(shmfd->shm_object, 0, 1769 objsize); 1770 } 1771 error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags, 1772 shmfd->shm_object, foff, writecnt, td); 1773 } 1774 if (error != 0) { 1775 if (writecnt) 1776 vm_pager_release_writecount(shmfd->shm_object, 0, 1777 objsize); 1778 vm_object_deallocate(shmfd->shm_object); 1779 } 1780 out: 1781 shm_rangelock_unlock(shmfd, rl_cookie); 1782 return (error); 1783 } 1784 1785 static int 1786 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, 1787 struct thread *td) 1788 { 1789 struct shmfd *shmfd; 1790 int error; 1791 1792 error = 0; 1793 shmfd = fp->f_data; 1794 mtx_lock(&shm_timestamp_lock); 1795 /* 1796 * SUSv4 says that x bits of permission need not be affected. 1797 * Be consistent with our shm_open there. 1798 */ 1799 #ifdef MAC 1800 error = mac_posixshm_check_setmode(active_cred, shmfd, mode); 1801 if (error != 0) 1802 goto out; 1803 #endif 1804 error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid, 1805 VADMIN, active_cred); 1806 if (error != 0) 1807 goto out; 1808 shmfd->shm_mode = mode & ACCESSPERMS; 1809 out: 1810 mtx_unlock(&shm_timestamp_lock); 1811 return (error); 1812 } 1813 1814 static int 1815 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred, 1816 struct thread *td) 1817 { 1818 struct shmfd *shmfd; 1819 int error; 1820 1821 error = 0; 1822 shmfd = fp->f_data; 1823 mtx_lock(&shm_timestamp_lock); 1824 #ifdef MAC 1825 error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid); 1826 if (error != 0) 1827 goto out; 1828 #endif 1829 if (uid == (uid_t)-1) 1830 uid = shmfd->shm_uid; 1831 if (gid == (gid_t)-1) 1832 gid = shmfd->shm_gid; 1833 if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) || 1834 (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) && 1835 (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN))) 1836 goto out; 1837 shmfd->shm_uid = uid; 1838 shmfd->shm_gid = gid; 1839 out: 1840 mtx_unlock(&shm_timestamp_lock); 1841 return (error); 1842 } 1843 1844 /* 1845 * Helper routines to allow the backing object of a shared memory file 1846 * descriptor to be mapped in the kernel. 1847 */ 1848 int 1849 shm_map(struct file *fp, size_t size, off_t offset, void **memp) 1850 { 1851 struct shmfd *shmfd; 1852 vm_offset_t kva, ofs; 1853 vm_object_t obj; 1854 int rv; 1855 1856 if (fp->f_type != DTYPE_SHM) 1857 return (EINVAL); 1858 shmfd = fp->f_data; 1859 obj = shmfd->shm_object; 1860 VM_OBJECT_WLOCK(obj); 1861 /* 1862 * XXXRW: This validation is probably insufficient, and subject to 1863 * sign errors. It should be fixed. 1864 */ 1865 if (offset >= shmfd->shm_size || 1866 offset + size > round_page(shmfd->shm_size)) { 1867 VM_OBJECT_WUNLOCK(obj); 1868 return (EINVAL); 1869 } 1870 1871 shmfd->shm_kmappings++; 1872 vm_object_reference_locked(obj); 1873 VM_OBJECT_WUNLOCK(obj); 1874 1875 /* Map the object into the kernel_map and wire it. */ 1876 kva = vm_map_min(kernel_map); 1877 ofs = offset & PAGE_MASK; 1878 offset = trunc_page(offset); 1879 size = round_page(size + ofs); 1880 rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0, 1881 VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE, 1882 VM_PROT_READ | VM_PROT_WRITE, 0); 1883 if (rv == KERN_SUCCESS) { 1884 rv = vm_map_wire(kernel_map, kva, kva + size, 1885 VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES); 1886 if (rv == KERN_SUCCESS) { 1887 *memp = (void *)(kva + ofs); 1888 return (0); 1889 } 1890 vm_map_remove(kernel_map, kva, kva + size); 1891 } else 1892 vm_object_deallocate(obj); 1893 1894 /* On failure, drop our mapping reference. */ 1895 VM_OBJECT_WLOCK(obj); 1896 shmfd->shm_kmappings--; 1897 VM_OBJECT_WUNLOCK(obj); 1898 1899 return (vm_mmap_to_errno(rv)); 1900 } 1901 1902 /* 1903 * We require the caller to unmap the entire entry. This allows us to 1904 * safely decrement shm_kmappings when a mapping is removed. 1905 */ 1906 int 1907 shm_unmap(struct file *fp, void *mem, size_t size) 1908 { 1909 struct shmfd *shmfd; 1910 vm_map_entry_t entry; 1911 vm_offset_t kva, ofs; 1912 vm_object_t obj; 1913 vm_pindex_t pindex; 1914 vm_prot_t prot; 1915 boolean_t wired; 1916 vm_map_t map; 1917 int rv; 1918 1919 if (fp->f_type != DTYPE_SHM) 1920 return (EINVAL); 1921 shmfd = fp->f_data; 1922 kva = (vm_offset_t)mem; 1923 ofs = kva & PAGE_MASK; 1924 kva = trunc_page(kva); 1925 size = round_page(size + ofs); 1926 map = kernel_map; 1927 rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry, 1928 &obj, &pindex, &prot, &wired); 1929 if (rv != KERN_SUCCESS) 1930 return (EINVAL); 1931 if (entry->start != kva || entry->end != kva + size) { 1932 vm_map_lookup_done(map, entry); 1933 return (EINVAL); 1934 } 1935 vm_map_lookup_done(map, entry); 1936 if (obj != shmfd->shm_object) 1937 return (EINVAL); 1938 vm_map_remove(map, kva, kva + size); 1939 VM_OBJECT_WLOCK(obj); 1940 KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped")); 1941 shmfd->shm_kmappings--; 1942 VM_OBJECT_WUNLOCK(obj); 1943 return (0); 1944 } 1945 1946 static int 1947 shm_fill_kinfo_locked(struct shmfd *shmfd, struct kinfo_file *kif, bool list) 1948 { 1949 const char *path, *pr_path; 1950 size_t pr_pathlen; 1951 bool visible; 1952 1953 sx_assert(&shm_dict_lock, SA_LOCKED); 1954 kif->kf_type = KF_TYPE_SHM; 1955 kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode; 1956 kif->kf_un.kf_file.kf_file_size = shmfd->shm_size; 1957 if (shmfd->shm_path != NULL) { 1958 path = shmfd->shm_path; 1959 pr_path = curthread->td_ucred->cr_prison->pr_path; 1960 if (strcmp(pr_path, "/") != 0) { 1961 /* Return the jail-rooted pathname. */ 1962 pr_pathlen = strlen(pr_path); 1963 visible = strncmp(path, pr_path, pr_pathlen) == 0 && 1964 path[pr_pathlen] == '/'; 1965 if (list && !visible) 1966 return (EPERM); 1967 if (visible) 1968 path += pr_pathlen; 1969 } 1970 strlcpy(kif->kf_path, path, sizeof(kif->kf_path)); 1971 } 1972 return (0); 1973 } 1974 1975 static int 1976 shm_fill_kinfo(struct file *fp, struct kinfo_file *kif, 1977 struct filedesc *fdp __unused) 1978 { 1979 int res; 1980 1981 sx_slock(&shm_dict_lock); 1982 res = shm_fill_kinfo_locked(fp->f_data, kif, false); 1983 sx_sunlock(&shm_dict_lock); 1984 return (res); 1985 } 1986 1987 static int 1988 shm_add_seals(struct file *fp, int seals) 1989 { 1990 struct shmfd *shmfd; 1991 void *rl_cookie; 1992 vm_ooffset_t writemappings; 1993 int error, nseals; 1994 1995 error = 0; 1996 shmfd = fp->f_data; 1997 rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX); 1998 1999 /* Even already-set seals should result in EPERM. */ 2000 if ((shmfd->shm_seals & F_SEAL_SEAL) != 0) { 2001 error = EPERM; 2002 goto out; 2003 } 2004 nseals = seals & ~shmfd->shm_seals; 2005 if ((nseals & F_SEAL_WRITE) != 0) { 2006 if (shm_largepage(shmfd)) { 2007 error = ENOTSUP; 2008 goto out; 2009 } 2010 2011 /* 2012 * The rangelock above prevents writable mappings from being 2013 * added after we've started applying seals. The RLOCK here 2014 * is to avoid torn reads on ILP32 arches as unmapping/reducing 2015 * writemappings will be done without a rangelock. 2016 */ 2017 VM_OBJECT_RLOCK(shmfd->shm_object); 2018 writemappings = shmfd->shm_object->un_pager.swp.writemappings; 2019 VM_OBJECT_RUNLOCK(shmfd->shm_object); 2020 /* kmappings are also writable */ 2021 if (writemappings > 0) { 2022 error = EBUSY; 2023 goto out; 2024 } 2025 } 2026 shmfd->shm_seals |= nseals; 2027 out: 2028 shm_rangelock_unlock(shmfd, rl_cookie); 2029 return (error); 2030 } 2031 2032 static int 2033 shm_get_seals(struct file *fp, int *seals) 2034 { 2035 struct shmfd *shmfd; 2036 2037 shmfd = fp->f_data; 2038 *seals = shmfd->shm_seals; 2039 return (0); 2040 } 2041 2042 static int 2043 shm_deallocate(struct shmfd *shmfd, off_t *offset, off_t *length, int flags) 2044 { 2045 vm_object_t object; 2046 vm_pindex_t pistart, pi, piend; 2047 vm_ooffset_t off, len; 2048 int startofs, endofs, end; 2049 int error; 2050 2051 off = *offset; 2052 len = *length; 2053 KASSERT(off + len <= (vm_ooffset_t)OFF_MAX, ("off + len overflows")); 2054 if (off + len > shmfd->shm_size) 2055 len = shmfd->shm_size - off; 2056 object = shmfd->shm_object; 2057 startofs = off & PAGE_MASK; 2058 endofs = (off + len) & PAGE_MASK; 2059 pistart = OFF_TO_IDX(off); 2060 piend = OFF_TO_IDX(off + len); 2061 pi = OFF_TO_IDX(off + PAGE_MASK); 2062 error = 0; 2063 2064 /* Handle the case when offset is on or beyond shm size. */ 2065 if ((off_t)len <= 0) { 2066 *length = 0; 2067 return (0); 2068 } 2069 2070 VM_OBJECT_WLOCK(object); 2071 2072 if (startofs != 0) { 2073 end = pistart != piend ? PAGE_SIZE : endofs; 2074 error = shm_partial_page_invalidate(object, pistart, startofs, 2075 end); 2076 if (error) 2077 goto out; 2078 off += end - startofs; 2079 len -= end - startofs; 2080 } 2081 2082 if (pi < piend) { 2083 vm_object_page_remove(object, pi, piend, 0); 2084 off += IDX_TO_OFF(piend - pi); 2085 len -= IDX_TO_OFF(piend - pi); 2086 } 2087 2088 if (endofs != 0 && pistart != piend) { 2089 error = shm_partial_page_invalidate(object, piend, 0, endofs); 2090 if (error) 2091 goto out; 2092 off += endofs; 2093 len -= endofs; 2094 } 2095 2096 out: 2097 VM_OBJECT_WUNLOCK(object); 2098 *offset = off; 2099 *length = len; 2100 return (error); 2101 } 2102 2103 static int 2104 shm_fspacectl(struct file *fp, int cmd, off_t *offset, off_t *length, int flags, 2105 struct ucred *active_cred, struct thread *td) 2106 { 2107 void *rl_cookie; 2108 struct shmfd *shmfd; 2109 off_t off, len; 2110 int error; 2111 2112 KASSERT(cmd == SPACECTL_DEALLOC, ("shm_fspacectl: Invalid cmd")); 2113 KASSERT((flags & ~SPACECTL_F_SUPPORTED) == 0, 2114 ("shm_fspacectl: non-zero flags")); 2115 KASSERT(*offset >= 0 && *length > 0 && *length <= OFF_MAX - *offset, 2116 ("shm_fspacectl: offset/length overflow or underflow")); 2117 2118 shmfd = fp->f_data; 2119 off = *offset; 2120 len = *length; 2121 2122 if (shm_largepage(shmfd)) 2123 return (ENOTSUP); 2124 2125 rl_cookie = shm_rangelock_wlock(shmfd, off, off + len); 2126 switch (cmd) { 2127 case SPACECTL_DEALLOC: 2128 if ((shmfd->shm_seals & F_SEAL_WRITE) != 0) { 2129 error = EPERM; 2130 break; 2131 } 2132 error = shm_deallocate(shmfd, &off, &len, flags); 2133 *offset = off; 2134 *length = len; 2135 break; 2136 default: 2137 __assert_unreachable(); 2138 } 2139 shm_rangelock_unlock(shmfd, rl_cookie); 2140 return (error); 2141 } 2142 2143 2144 static int 2145 shm_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td) 2146 { 2147 void *rl_cookie; 2148 struct shmfd *shmfd; 2149 size_t size; 2150 int error; 2151 2152 /* This assumes that the caller already checked for overflow. */ 2153 error = 0; 2154 shmfd = fp->f_data; 2155 size = offset + len; 2156 2157 /* 2158 * Just grab the rangelock for the range that we may be attempting to 2159 * grow, rather than blocking read/write for regions we won't be 2160 * touching while this (potential) resize is in progress. Other 2161 * attempts to resize the shmfd will have to take a write lock from 0 to 2162 * OFF_MAX, so this being potentially beyond the current usable range of 2163 * the shmfd is not necessarily a concern. If other mechanisms are 2164 * added to grow a shmfd, this may need to be re-evaluated. 2165 */ 2166 rl_cookie = shm_rangelock_wlock(shmfd, offset, size); 2167 if (size > shmfd->shm_size) 2168 error = shm_dotruncate_cookie(shmfd, size, rl_cookie); 2169 shm_rangelock_unlock(shmfd, rl_cookie); 2170 /* Translate to posix_fallocate(2) return value as needed. */ 2171 if (error == ENOMEM) 2172 error = ENOSPC; 2173 return (error); 2174 } 2175 2176 static int 2177 sysctl_posix_shm_list(SYSCTL_HANDLER_ARGS) 2178 { 2179 struct shm_mapping *shmm; 2180 struct sbuf sb; 2181 struct kinfo_file kif = {}; 2182 u_long i; 2183 int error, error2; 2184 2185 sbuf_new_for_sysctl(&sb, NULL, sizeof(struct kinfo_file) * 5, req); 2186 sbuf_clear_flags(&sb, SBUF_INCLUDENUL); 2187 error = 0; 2188 sx_slock(&shm_dict_lock); 2189 for (i = 0; i < shm_hash + 1; i++) { 2190 LIST_FOREACH(shmm, &shm_dictionary[i], sm_link) { 2191 error = shm_fill_kinfo_locked(shmm->sm_shmfd, 2192 &kif, true); 2193 if (error == EPERM) { 2194 error = 0; 2195 continue; 2196 } 2197 if (error != 0) 2198 break; 2199 pack_kinfo(&kif); 2200 error = sbuf_bcat(&sb, &kif, kif.kf_structsize) == 0 ? 2201 0 : ENOMEM; 2202 if (error != 0) 2203 break; 2204 } 2205 } 2206 sx_sunlock(&shm_dict_lock); 2207 error2 = sbuf_finish(&sb); 2208 sbuf_delete(&sb); 2209 return (error != 0 ? error : error2); 2210 } 2211 2212 SYSCTL_PROC(_kern_ipc, OID_AUTO, posix_shm_list, 2213 CTLFLAG_RD | CTLFLAG_PRISON | CTLFLAG_MPSAFE | CTLTYPE_OPAQUE, 2214 NULL, 0, sysctl_posix_shm_list, "", 2215 "POSIX SHM list"); 2216 2217 int 2218 kern_shm_open(struct thread *td, const char *path, int flags, mode_t mode, 2219 struct filecaps *caps) 2220 { 2221 2222 return (kern_shm_open2(td, path, flags, mode, 0, caps, NULL, NULL)); 2223 } 2224 2225 /* 2226 * This version of the shm_open() interface leaves CLOEXEC behavior up to the 2227 * caller, and libc will enforce it for the traditional shm_open() call. This 2228 * allows other consumers, like memfd_create(), to opt-in for CLOEXEC. This 2229 * interface also includes a 'name' argument that is currently unused, but could 2230 * potentially be exported later via some interface for debugging purposes. 2231 * From the kernel's perspective, it is optional. Individual consumers like 2232 * memfd_create() may require it in order to be compatible with other systems 2233 * implementing the same function. 2234 */ 2235 int 2236 sys_shm_open2(struct thread *td, struct shm_open2_args *uap) 2237 { 2238 2239 return (kern_shm_open2(td, uap->path, uap->flags, uap->mode, 2240 uap->shmflags, NULL, uap->name, NULL)); 2241 } 2242 2243 int 2244 shm_get_path(struct vm_object *obj, char *path, size_t sz) 2245 { 2246 struct shmfd *shmfd; 2247 int error; 2248 2249 error = 0; 2250 shmfd = NULL; 2251 sx_slock(&shm_dict_lock); 2252 VM_OBJECT_RLOCK(obj); 2253 if ((obj->flags & OBJ_POSIXSHM) == 0) { 2254 error = EINVAL; 2255 } else { 2256 if (obj->type == shmfd_pager_type) 2257 shmfd = obj->un_pager.swp.swp_priv; 2258 else if (obj->type == OBJT_PHYS) 2259 shmfd = obj->un_pager.phys.phys_priv; 2260 if (shmfd == NULL) { 2261 error = ENXIO; 2262 } else { 2263 strlcpy(path, shmfd->shm_path == NULL ? "anon" : 2264 shmfd->shm_path, sz); 2265 } 2266 } 2267 if (error != 0) 2268 path[0] = '\0'; 2269 VM_OBJECT_RUNLOCK(obj); 2270 sx_sunlock(&shm_dict_lock); 2271 return (error); 2272 } 2273