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)) 188 #define shm_rangelock_rlock(shmfd, start, end) \ 189 rangelock_rlock(&(shmfd)->shm_rl, (start), (end)) 190 #define shm_rangelock_tryrlock(shmfd, start, end) \ 191 rangelock_tryrlock(&(shmfd)->shm_rl, (start), (end)) 192 #define shm_rangelock_wlock(shmfd, start, end) \ 193 rangelock_wlock(&(shmfd)->shm_rl, (start), (end)) 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 obj = shmfd->shm_object = phys_pager_allocate(NULL, 949 &shm_largepage_phys_ops, NULL, shmfd->shm_size, 950 VM_PROT_DEFAULT, 0, ucred); 951 VM_OBJECT_WLOCK(shmfd->shm_object); 952 obj->un_pager.phys.phys_priv = shmfd; 953 vm_object_set_flag(obj, OBJ_POSIXSHM); 954 VM_OBJECT_WUNLOCK(shmfd->shm_object); 955 shmfd->shm_lp_alloc_policy = SHM_LARGEPAGE_ALLOC_DEFAULT; 956 } else { 957 obj = vm_pager_allocate(shmfd_pager_type, NULL, 958 shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred); 959 VM_OBJECT_WLOCK(obj); 960 obj->un_pager.swp.swp_priv = shmfd; 961 vm_object_set_flag(obj, OBJ_POSIXSHM); 962 VM_OBJECT_WUNLOCK(obj); 963 shmfd->shm_object = obj; 964 } 965 KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate")); 966 vfs_timestamp(&shmfd->shm_birthtime); 967 shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime = 968 shmfd->shm_birthtime; 969 shmfd->shm_ino = alloc_unr64(&shm_ino_unr); 970 refcount_init(&shmfd->shm_refs, 1); 971 mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF); 972 rangelock_init(&shmfd->shm_rl); 973 #ifdef MAC 974 mac_posixshm_init(shmfd); 975 mac_posixshm_create(ucred, shmfd); 976 #endif 977 978 return (shmfd); 979 } 980 981 struct shmfd * 982 shm_hold(struct shmfd *shmfd) 983 { 984 985 refcount_acquire(&shmfd->shm_refs); 986 return (shmfd); 987 } 988 989 void 990 shm_drop(struct shmfd *shmfd) 991 { 992 vm_object_t obj; 993 994 if (refcount_release(&shmfd->shm_refs)) { 995 #ifdef MAC 996 mac_posixshm_destroy(shmfd); 997 #endif 998 rangelock_destroy(&shmfd->shm_rl); 999 mtx_destroy(&shmfd->shm_mtx); 1000 obj = shmfd->shm_object; 1001 VM_OBJECT_WLOCK(obj); 1002 if (shm_largepage(shmfd)) 1003 obj->un_pager.phys.phys_priv = NULL; 1004 else 1005 obj->un_pager.swp.swp_priv = NULL; 1006 VM_OBJECT_WUNLOCK(obj); 1007 vm_object_deallocate(obj); 1008 free(shmfd, M_SHMFD); 1009 } 1010 } 1011 1012 /* 1013 * Determine if the credentials have sufficient permissions for a 1014 * specified combination of FREAD and FWRITE. 1015 */ 1016 int 1017 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags) 1018 { 1019 accmode_t accmode; 1020 int error; 1021 1022 accmode = 0; 1023 if (flags & FREAD) 1024 accmode |= VREAD; 1025 if (flags & FWRITE) 1026 accmode |= VWRITE; 1027 mtx_lock(&shm_timestamp_lock); 1028 error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid, 1029 accmode, ucred); 1030 mtx_unlock(&shm_timestamp_lock); 1031 return (error); 1032 } 1033 1034 static void 1035 shm_init(void *arg) 1036 { 1037 char name[32]; 1038 int i; 1039 1040 mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF); 1041 sx_init(&shm_dict_lock, "shm dictionary"); 1042 shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash); 1043 new_unrhdr64(&shm_ino_unr, 1); 1044 shm_dev_ino = devfs_alloc_cdp_inode(); 1045 KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized")); 1046 shmfd_pager_type = vm_pager_alloc_dyn_type(&shm_swap_pager_ops, 1047 OBJT_SWAP); 1048 MPASS(shmfd_pager_type != -1); 1049 1050 for (i = 1; i < MAXPAGESIZES; i++) { 1051 if (pagesizes[i] == 0) 1052 break; 1053 #define M (1024 * 1024) 1054 #define G (1024 * M) 1055 if (pagesizes[i] >= G) 1056 snprintf(name, sizeof(name), "%luG", pagesizes[i] / G); 1057 else if (pagesizes[i] >= M) 1058 snprintf(name, sizeof(name), "%luM", pagesizes[i] / M); 1059 else 1060 snprintf(name, sizeof(name), "%lu", pagesizes[i]); 1061 #undef G 1062 #undef M 1063 SYSCTL_ADD_ULONG(NULL, SYSCTL_STATIC_CHILDREN(_vm_largepages), 1064 OID_AUTO, name, CTLFLAG_RD, &count_largepages[i], 1065 "number of non-transient largepages allocated"); 1066 } 1067 } 1068 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL); 1069 1070 /* 1071 * Remove all shared memory objects that belong to a prison. 1072 */ 1073 void 1074 shm_remove_prison(struct prison *pr) 1075 { 1076 struct shm_mapping *shmm, *tshmm; 1077 u_long i; 1078 1079 sx_xlock(&shm_dict_lock); 1080 for (i = 0; i < shm_hash + 1; i++) { 1081 LIST_FOREACH_SAFE(shmm, &shm_dictionary[i], sm_link, tshmm) { 1082 if (shmm->sm_shmfd->shm_object->cred && 1083 shmm->sm_shmfd->shm_object->cred->cr_prison == pr) 1084 shm_doremove(shmm); 1085 } 1086 } 1087 sx_xunlock(&shm_dict_lock); 1088 } 1089 1090 /* 1091 * Dictionary management. We maintain an in-kernel dictionary to map 1092 * paths to shmfd objects. We use the FNV hash on the path to store 1093 * the mappings in a hash table. 1094 */ 1095 static struct shmfd * 1096 shm_lookup(char *path, Fnv32_t fnv) 1097 { 1098 struct shm_mapping *map; 1099 1100 LIST_FOREACH(map, SHM_HASH(fnv), sm_link) { 1101 if (map->sm_fnv != fnv) 1102 continue; 1103 if (strcmp(map->sm_path, path) == 0) 1104 return (map->sm_shmfd); 1105 } 1106 1107 return (NULL); 1108 } 1109 1110 static void 1111 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd) 1112 { 1113 struct shm_mapping *map; 1114 1115 map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK); 1116 map->sm_path = path; 1117 map->sm_fnv = fnv; 1118 map->sm_shmfd = shm_hold(shmfd); 1119 shmfd->shm_path = path; 1120 LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link); 1121 } 1122 1123 static int 1124 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred) 1125 { 1126 struct shm_mapping *map; 1127 int error; 1128 1129 LIST_FOREACH(map, SHM_HASH(fnv), sm_link) { 1130 if (map->sm_fnv != fnv) 1131 continue; 1132 if (strcmp(map->sm_path, path) == 0) { 1133 #ifdef MAC 1134 error = mac_posixshm_check_unlink(ucred, map->sm_shmfd); 1135 if (error) 1136 return (error); 1137 #endif 1138 error = shm_access(map->sm_shmfd, ucred, 1139 FREAD | FWRITE); 1140 if (error) 1141 return (error); 1142 shm_doremove(map); 1143 return (0); 1144 } 1145 } 1146 1147 return (ENOENT); 1148 } 1149 1150 static void 1151 shm_doremove(struct shm_mapping *map) 1152 { 1153 map->sm_shmfd->shm_path = NULL; 1154 LIST_REMOVE(map, sm_link); 1155 shm_drop(map->sm_shmfd); 1156 free(map->sm_path, M_SHMFD); 1157 free(map, M_SHMFD); 1158 } 1159 1160 int 1161 kern_shm_open2(struct thread *td, const char *userpath, int flags, mode_t mode, 1162 int shmflags, struct filecaps *fcaps, const char *name __unused) 1163 { 1164 struct pwddesc *pdp; 1165 struct shmfd *shmfd; 1166 struct file *fp; 1167 char *path; 1168 void *rl_cookie; 1169 Fnv32_t fnv; 1170 mode_t cmode; 1171 int error, fd, initial_seals; 1172 bool largepage; 1173 1174 if ((shmflags & ~(SHM_ALLOW_SEALING | SHM_GROW_ON_WRITE | 1175 SHM_LARGEPAGE)) != 0) 1176 return (EINVAL); 1177 1178 initial_seals = F_SEAL_SEAL; 1179 if ((shmflags & SHM_ALLOW_SEALING) != 0) 1180 initial_seals &= ~F_SEAL_SEAL; 1181 1182 AUDIT_ARG_FFLAGS(flags); 1183 AUDIT_ARG_MODE(mode); 1184 1185 if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR) 1186 return (EINVAL); 1187 1188 if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 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 free(path, M_SHMFD); 1219 return (ECAPMODE); 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) { 1240 free(path, M_SHMFD); 1241 return (error); 1242 } 1243 1244 /* A SHM_ANON path pointer creates an anonymous object. */ 1245 if (userpath == SHM_ANON) { 1246 /* A read-only anonymous object is pointless. */ 1247 if ((flags & O_ACCMODE) == O_RDONLY) { 1248 fdclose(td, fp, fd); 1249 fdrop(fp, td); 1250 return (EINVAL); 1251 } 1252 shmfd = shm_alloc(td->td_ucred, cmode, largepage); 1253 shmfd->shm_seals = initial_seals; 1254 shmfd->shm_flags = shmflags; 1255 } else { 1256 fnv = fnv_32_str(path, FNV1_32_INIT); 1257 sx_xlock(&shm_dict_lock); 1258 shmfd = shm_lookup(path, fnv); 1259 if (shmfd == NULL) { 1260 /* Object does not yet exist, create it if requested. */ 1261 if (flags & O_CREAT) { 1262 #ifdef MAC 1263 error = mac_posixshm_check_create(td->td_ucred, 1264 path); 1265 if (error == 0) { 1266 #endif 1267 shmfd = shm_alloc(td->td_ucred, cmode, 1268 largepage); 1269 shmfd->shm_seals = initial_seals; 1270 shmfd->shm_flags = shmflags; 1271 shm_insert(path, fnv, shmfd); 1272 #ifdef MAC 1273 } 1274 #endif 1275 } else { 1276 free(path, M_SHMFD); 1277 error = ENOENT; 1278 } 1279 } else { 1280 rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX); 1281 1282 /* 1283 * kern_shm_open() likely shouldn't ever error out on 1284 * trying to set a seal that already exists, unlike 1285 * F_ADD_SEALS. This would break terribly as 1286 * shm_open(2) actually sets F_SEAL_SEAL to maintain 1287 * historical behavior where the underlying file could 1288 * not be sealed. 1289 */ 1290 initial_seals &= ~shmfd->shm_seals; 1291 1292 /* 1293 * Object already exists, obtain a new 1294 * reference if requested and permitted. 1295 */ 1296 free(path, M_SHMFD); 1297 1298 /* 1299 * initial_seals can't set additional seals if we've 1300 * already been set F_SEAL_SEAL. If F_SEAL_SEAL is set, 1301 * then we've already removed that one from 1302 * initial_seals. This is currently redundant as we 1303 * only allow setting F_SEAL_SEAL at creation time, but 1304 * it's cheap to check and decreases the effort required 1305 * to allow additional seals. 1306 */ 1307 if ((shmfd->shm_seals & F_SEAL_SEAL) != 0 && 1308 initial_seals != 0) 1309 error = EPERM; 1310 else if ((flags & (O_CREAT | O_EXCL)) == 1311 (O_CREAT | O_EXCL)) 1312 error = EEXIST; 1313 else if (shmflags != 0 && shmflags != shmfd->shm_flags) 1314 error = EINVAL; 1315 else { 1316 #ifdef MAC 1317 error = mac_posixshm_check_open(td->td_ucred, 1318 shmfd, FFLAGS(flags & O_ACCMODE)); 1319 if (error == 0) 1320 #endif 1321 error = shm_access(shmfd, td->td_ucred, 1322 FFLAGS(flags & O_ACCMODE)); 1323 } 1324 1325 /* 1326 * Truncate the file back to zero length if 1327 * O_TRUNC was specified and the object was 1328 * opened with read/write. 1329 */ 1330 if (error == 0 && 1331 (flags & (O_ACCMODE | O_TRUNC)) == 1332 (O_RDWR | O_TRUNC)) { 1333 VM_OBJECT_WLOCK(shmfd->shm_object); 1334 #ifdef MAC 1335 error = mac_posixshm_check_truncate( 1336 td->td_ucred, fp->f_cred, shmfd); 1337 if (error == 0) 1338 #endif 1339 error = shm_dotruncate_locked(shmfd, 0, 1340 rl_cookie); 1341 VM_OBJECT_WUNLOCK(shmfd->shm_object); 1342 } 1343 if (error == 0) { 1344 /* 1345 * Currently we only allow F_SEAL_SEAL to be 1346 * set initially. As noted above, this would 1347 * need to be reworked should that change. 1348 */ 1349 shmfd->shm_seals |= initial_seals; 1350 shm_hold(shmfd); 1351 } 1352 shm_rangelock_unlock(shmfd, rl_cookie); 1353 } 1354 sx_xunlock(&shm_dict_lock); 1355 1356 if (error) { 1357 fdclose(td, fp, fd); 1358 fdrop(fp, td); 1359 return (error); 1360 } 1361 } 1362 1363 finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops); 1364 1365 td->td_retval[0] = fd; 1366 fdrop(fp, td); 1367 1368 return (0); 1369 } 1370 1371 /* System calls. */ 1372 #ifdef COMPAT_FREEBSD12 1373 int 1374 freebsd12_shm_open(struct thread *td, struct freebsd12_shm_open_args *uap) 1375 { 1376 1377 return (kern_shm_open(td, uap->path, uap->flags | O_CLOEXEC, 1378 uap->mode, NULL)); 1379 } 1380 #endif 1381 1382 int 1383 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap) 1384 { 1385 char *path; 1386 Fnv32_t fnv; 1387 int error; 1388 1389 error = shm_copyin_path(td, uap->path, &path); 1390 if (error != 0) 1391 return (error); 1392 1393 AUDIT_ARG_UPATH1_CANON(path); 1394 fnv = fnv_32_str(path, FNV1_32_INIT); 1395 sx_xlock(&shm_dict_lock); 1396 error = shm_remove(path, fnv, td->td_ucred); 1397 sx_xunlock(&shm_dict_lock); 1398 free(path, M_SHMFD); 1399 1400 return (error); 1401 } 1402 1403 int 1404 sys_shm_rename(struct thread *td, struct shm_rename_args *uap) 1405 { 1406 char *path_from = NULL, *path_to = NULL; 1407 Fnv32_t fnv_from, fnv_to; 1408 struct shmfd *fd_from; 1409 struct shmfd *fd_to; 1410 int error; 1411 int flags; 1412 1413 flags = uap->flags; 1414 AUDIT_ARG_FFLAGS(flags); 1415 1416 /* 1417 * Make sure the user passed only valid flags. 1418 * If you add a new flag, please add a new term here. 1419 */ 1420 if ((flags & ~( 1421 SHM_RENAME_NOREPLACE | 1422 SHM_RENAME_EXCHANGE 1423 )) != 0) { 1424 error = EINVAL; 1425 goto out; 1426 } 1427 1428 /* 1429 * EXCHANGE and NOREPLACE don't quite make sense together. Let's 1430 * force the user to choose one or the other. 1431 */ 1432 if ((flags & SHM_RENAME_NOREPLACE) != 0 && 1433 (flags & SHM_RENAME_EXCHANGE) != 0) { 1434 error = EINVAL; 1435 goto out; 1436 } 1437 1438 /* Renaming to or from anonymous makes no sense */ 1439 if (uap->path_from == SHM_ANON || uap->path_to == SHM_ANON) { 1440 error = EINVAL; 1441 goto out; 1442 } 1443 1444 error = shm_copyin_path(td, uap->path_from, &path_from); 1445 if (error != 0) 1446 goto out; 1447 1448 error = shm_copyin_path(td, uap->path_to, &path_to); 1449 if (error != 0) 1450 goto out; 1451 1452 AUDIT_ARG_UPATH1_CANON(path_from); 1453 AUDIT_ARG_UPATH2_CANON(path_to); 1454 1455 /* Rename with from/to equal is a no-op */ 1456 if (strcmp(path_from, path_to) == 0) 1457 goto out; 1458 1459 fnv_from = fnv_32_str(path_from, FNV1_32_INIT); 1460 fnv_to = fnv_32_str(path_to, FNV1_32_INIT); 1461 1462 sx_xlock(&shm_dict_lock); 1463 1464 fd_from = shm_lookup(path_from, fnv_from); 1465 if (fd_from == NULL) { 1466 error = ENOENT; 1467 goto out_locked; 1468 } 1469 1470 fd_to = shm_lookup(path_to, fnv_to); 1471 if ((flags & SHM_RENAME_NOREPLACE) != 0 && fd_to != NULL) { 1472 error = EEXIST; 1473 goto out_locked; 1474 } 1475 1476 /* 1477 * Unconditionally prevents shm_remove from invalidating the 'from' 1478 * shm's state. 1479 */ 1480 shm_hold(fd_from); 1481 error = shm_remove(path_from, fnv_from, td->td_ucred); 1482 1483 /* 1484 * One of my assumptions failed if ENOENT (e.g. locking didn't 1485 * protect us) 1486 */ 1487 KASSERT(error != ENOENT, ("Our shm disappeared during shm_rename: %s", 1488 path_from)); 1489 if (error != 0) { 1490 shm_drop(fd_from); 1491 goto out_locked; 1492 } 1493 1494 /* 1495 * If we are exchanging, we need to ensure the shm_remove below 1496 * doesn't invalidate the dest shm's state. 1497 */ 1498 if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) 1499 shm_hold(fd_to); 1500 1501 /* 1502 * NOTE: if path_to is not already in the hash, c'est la vie; 1503 * it simply means we have nothing already at path_to to unlink. 1504 * That is the ENOENT case. 1505 * 1506 * If we somehow don't have access to unlink this guy, but 1507 * did for the shm at path_from, then relink the shm to path_from 1508 * and abort with EACCES. 1509 * 1510 * All other errors: that is weird; let's relink and abort the 1511 * operation. 1512 */ 1513 error = shm_remove(path_to, fnv_to, td->td_ucred); 1514 if (error != 0 && error != ENOENT) { 1515 shm_insert(path_from, fnv_from, fd_from); 1516 shm_drop(fd_from); 1517 /* Don't free path_from now, since the hash references it */ 1518 path_from = NULL; 1519 goto out_locked; 1520 } 1521 1522 error = 0; 1523 1524 shm_insert(path_to, fnv_to, fd_from); 1525 1526 /* Don't free path_to now, since the hash references it */ 1527 path_to = NULL; 1528 1529 /* We kept a ref when we removed, and incremented again in insert */ 1530 shm_drop(fd_from); 1531 KASSERT(fd_from->shm_refs > 0, ("Expected >0 refs; got: %d\n", 1532 fd_from->shm_refs)); 1533 1534 if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) { 1535 shm_insert(path_from, fnv_from, fd_to); 1536 path_from = NULL; 1537 shm_drop(fd_to); 1538 KASSERT(fd_to->shm_refs > 0, ("Expected >0 refs; got: %d\n", 1539 fd_to->shm_refs)); 1540 } 1541 1542 out_locked: 1543 sx_xunlock(&shm_dict_lock); 1544 1545 out: 1546 free(path_from, M_SHMFD); 1547 free(path_to, M_SHMFD); 1548 return (error); 1549 } 1550 1551 static int 1552 shm_mmap_large(struct shmfd *shmfd, vm_map_t map, vm_offset_t *addr, 1553 vm_size_t size, vm_prot_t prot, vm_prot_t max_prot, int flags, 1554 vm_ooffset_t foff, struct thread *td) 1555 { 1556 struct vmspace *vms; 1557 vm_map_entry_t next_entry, prev_entry; 1558 vm_offset_t align, mask, maxaddr; 1559 int docow, error, rv, try; 1560 bool curmap; 1561 1562 if (shmfd->shm_lp_psind == 0) 1563 return (EINVAL); 1564 1565 /* MAP_PRIVATE is disabled */ 1566 if ((flags & ~(MAP_SHARED | MAP_FIXED | MAP_EXCL | 1567 MAP_NOCORE | MAP_32BIT | MAP_ALIGNMENT_MASK)) != 0) 1568 return (EINVAL); 1569 1570 vms = td->td_proc->p_vmspace; 1571 curmap = map == &vms->vm_map; 1572 if (curmap) { 1573 error = kern_mmap_racct_check(td, map, size); 1574 if (error != 0) 1575 return (error); 1576 } 1577 1578 docow = shmfd->shm_lp_psind << MAP_SPLIT_BOUNDARY_SHIFT; 1579 docow |= MAP_INHERIT_SHARE; 1580 if ((flags & MAP_NOCORE) != 0) 1581 docow |= MAP_DISABLE_COREDUMP; 1582 1583 mask = pagesizes[shmfd->shm_lp_psind] - 1; 1584 if ((foff & mask) != 0) 1585 return (EINVAL); 1586 maxaddr = vm_map_max(map); 1587 if ((flags & MAP_32BIT) != 0 && maxaddr > MAP_32BIT_MAX_ADDR) 1588 maxaddr = MAP_32BIT_MAX_ADDR; 1589 if (size == 0 || (size & mask) != 0 || 1590 (*addr != 0 && ((*addr & mask) != 0 || 1591 *addr + size < *addr || *addr + size > maxaddr))) 1592 return (EINVAL); 1593 1594 align = flags & MAP_ALIGNMENT_MASK; 1595 if (align == 0) { 1596 align = pagesizes[shmfd->shm_lp_psind]; 1597 } else if (align == MAP_ALIGNED_SUPER) { 1598 /* 1599 * MAP_ALIGNED_SUPER is only supported on superpage sizes, 1600 * i.e., [1, VM_NRESERVLEVEL]. shmfd->shm_lp_psind < 1 is 1601 * handled above. 1602 */ 1603 if ( 1604 #if VM_NRESERVLEVEL > 0 1605 shmfd->shm_lp_psind > VM_NRESERVLEVEL 1606 #else 1607 shmfd->shm_lp_psind > 1 1608 #endif 1609 ) 1610 return (EINVAL); 1611 align = pagesizes[shmfd->shm_lp_psind]; 1612 } else { 1613 align >>= MAP_ALIGNMENT_SHIFT; 1614 align = 1ULL << align; 1615 /* Also handles overflow. */ 1616 if (align < pagesizes[shmfd->shm_lp_psind]) 1617 return (EINVAL); 1618 } 1619 1620 vm_map_lock(map); 1621 if ((flags & MAP_FIXED) == 0) { 1622 try = 1; 1623 if (curmap && (*addr == 0 || 1624 (*addr >= round_page((vm_offset_t)vms->vm_taddr) && 1625 *addr < round_page((vm_offset_t)vms->vm_daddr + 1626 lim_max(td, RLIMIT_DATA))))) { 1627 *addr = roundup2((vm_offset_t)vms->vm_daddr + 1628 lim_max(td, RLIMIT_DATA), 1629 pagesizes[shmfd->shm_lp_psind]); 1630 } 1631 again: 1632 rv = vm_map_find_aligned(map, addr, size, maxaddr, align); 1633 if (rv != KERN_SUCCESS) { 1634 if (try == 1) { 1635 try = 2; 1636 *addr = vm_map_min(map); 1637 if ((*addr & mask) != 0) 1638 *addr = (*addr + mask) & mask; 1639 goto again; 1640 } 1641 goto fail1; 1642 } 1643 } else if ((flags & MAP_EXCL) == 0) { 1644 rv = vm_map_delete(map, *addr, *addr + size); 1645 if (rv != KERN_SUCCESS) 1646 goto fail1; 1647 } else { 1648 error = ENOSPC; 1649 if (vm_map_lookup_entry(map, *addr, &prev_entry)) 1650 goto fail; 1651 next_entry = vm_map_entry_succ(prev_entry); 1652 if (next_entry->start < *addr + size) 1653 goto fail; 1654 } 1655 1656 rv = vm_map_insert(map, shmfd->shm_object, foff, *addr, *addr + size, 1657 prot, max_prot, docow); 1658 fail1: 1659 error = vm_mmap_to_errno(rv); 1660 fail: 1661 vm_map_unlock(map); 1662 return (error); 1663 } 1664 1665 static int 1666 shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize, 1667 vm_prot_t prot, vm_prot_t max_maxprot, int flags, 1668 vm_ooffset_t foff, struct thread *td) 1669 { 1670 struct shmfd *shmfd; 1671 vm_prot_t maxprot; 1672 int error; 1673 bool writecnt; 1674 void *rl_cookie; 1675 1676 shmfd = fp->f_data; 1677 maxprot = VM_PROT_NONE; 1678 1679 rl_cookie = shm_rangelock_rlock(shmfd, 0, objsize); 1680 /* FREAD should always be set. */ 1681 if ((fp->f_flag & FREAD) != 0) 1682 maxprot |= VM_PROT_EXECUTE | VM_PROT_READ; 1683 1684 /* 1685 * If FWRITE's set, we can allow VM_PROT_WRITE unless it's a shared 1686 * mapping with a write seal applied. Private mappings are always 1687 * writeable. 1688 */ 1689 if ((flags & MAP_SHARED) == 0) { 1690 if ((max_maxprot & VM_PROT_WRITE) != 0) 1691 maxprot |= VM_PROT_WRITE; 1692 writecnt = false; 1693 } else { 1694 if ((fp->f_flag & FWRITE) != 0 && 1695 (shmfd->shm_seals & F_SEAL_WRITE) == 0) 1696 maxprot |= VM_PROT_WRITE; 1697 1698 /* 1699 * Any mappings from a writable descriptor may be upgraded to 1700 * VM_PROT_WRITE with mprotect(2), unless a write-seal was 1701 * applied between the open and subsequent mmap(2). We want to 1702 * reject application of a write seal as long as any such 1703 * mapping exists so that the seal cannot be trivially bypassed. 1704 */ 1705 writecnt = (maxprot & VM_PROT_WRITE) != 0; 1706 if (!writecnt && (prot & VM_PROT_WRITE) != 0) { 1707 error = EACCES; 1708 goto out; 1709 } 1710 } 1711 maxprot &= max_maxprot; 1712 1713 /* See comment in vn_mmap(). */ 1714 if ( 1715 #ifdef _LP64 1716 objsize > OFF_MAX || 1717 #endif 1718 foff > OFF_MAX - objsize) { 1719 error = EINVAL; 1720 goto out; 1721 } 1722 1723 #ifdef MAC 1724 error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags); 1725 if (error != 0) 1726 goto out; 1727 #endif 1728 1729 mtx_lock(&shm_timestamp_lock); 1730 vfs_timestamp(&shmfd->shm_atime); 1731 mtx_unlock(&shm_timestamp_lock); 1732 vm_object_reference(shmfd->shm_object); 1733 1734 if (shm_largepage(shmfd)) { 1735 writecnt = false; 1736 error = shm_mmap_large(shmfd, map, addr, objsize, prot, 1737 maxprot, flags, foff, td); 1738 } else { 1739 if (writecnt) { 1740 vm_pager_update_writecount(shmfd->shm_object, 0, 1741 objsize); 1742 } 1743 error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags, 1744 shmfd->shm_object, foff, writecnt, td); 1745 } 1746 if (error != 0) { 1747 if (writecnt) 1748 vm_pager_release_writecount(shmfd->shm_object, 0, 1749 objsize); 1750 vm_object_deallocate(shmfd->shm_object); 1751 } 1752 out: 1753 shm_rangelock_unlock(shmfd, rl_cookie); 1754 return (error); 1755 } 1756 1757 static int 1758 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, 1759 struct thread *td) 1760 { 1761 struct shmfd *shmfd; 1762 int error; 1763 1764 error = 0; 1765 shmfd = fp->f_data; 1766 mtx_lock(&shm_timestamp_lock); 1767 /* 1768 * SUSv4 says that x bits of permission need not be affected. 1769 * Be consistent with our shm_open there. 1770 */ 1771 #ifdef MAC 1772 error = mac_posixshm_check_setmode(active_cred, shmfd, mode); 1773 if (error != 0) 1774 goto out; 1775 #endif 1776 error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid, 1777 VADMIN, active_cred); 1778 if (error != 0) 1779 goto out; 1780 shmfd->shm_mode = mode & ACCESSPERMS; 1781 out: 1782 mtx_unlock(&shm_timestamp_lock); 1783 return (error); 1784 } 1785 1786 static int 1787 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred, 1788 struct thread *td) 1789 { 1790 struct shmfd *shmfd; 1791 int error; 1792 1793 error = 0; 1794 shmfd = fp->f_data; 1795 mtx_lock(&shm_timestamp_lock); 1796 #ifdef MAC 1797 error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid); 1798 if (error != 0) 1799 goto out; 1800 #endif 1801 if (uid == (uid_t)-1) 1802 uid = shmfd->shm_uid; 1803 if (gid == (gid_t)-1) 1804 gid = shmfd->shm_gid; 1805 if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) || 1806 (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) && 1807 (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN))) 1808 goto out; 1809 shmfd->shm_uid = uid; 1810 shmfd->shm_gid = gid; 1811 out: 1812 mtx_unlock(&shm_timestamp_lock); 1813 return (error); 1814 } 1815 1816 /* 1817 * Helper routines to allow the backing object of a shared memory file 1818 * descriptor to be mapped in the kernel. 1819 */ 1820 int 1821 shm_map(struct file *fp, size_t size, off_t offset, void **memp) 1822 { 1823 struct shmfd *shmfd; 1824 vm_offset_t kva, ofs; 1825 vm_object_t obj; 1826 int rv; 1827 1828 if (fp->f_type != DTYPE_SHM) 1829 return (EINVAL); 1830 shmfd = fp->f_data; 1831 obj = shmfd->shm_object; 1832 VM_OBJECT_WLOCK(obj); 1833 /* 1834 * XXXRW: This validation is probably insufficient, and subject to 1835 * sign errors. It should be fixed. 1836 */ 1837 if (offset >= shmfd->shm_size || 1838 offset + size > round_page(shmfd->shm_size)) { 1839 VM_OBJECT_WUNLOCK(obj); 1840 return (EINVAL); 1841 } 1842 1843 shmfd->shm_kmappings++; 1844 vm_object_reference_locked(obj); 1845 VM_OBJECT_WUNLOCK(obj); 1846 1847 /* Map the object into the kernel_map and wire it. */ 1848 kva = vm_map_min(kernel_map); 1849 ofs = offset & PAGE_MASK; 1850 offset = trunc_page(offset); 1851 size = round_page(size + ofs); 1852 rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0, 1853 VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE, 1854 VM_PROT_READ | VM_PROT_WRITE, 0); 1855 if (rv == KERN_SUCCESS) { 1856 rv = vm_map_wire(kernel_map, kva, kva + size, 1857 VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES); 1858 if (rv == KERN_SUCCESS) { 1859 *memp = (void *)(kva + ofs); 1860 return (0); 1861 } 1862 vm_map_remove(kernel_map, kva, kva + size); 1863 } else 1864 vm_object_deallocate(obj); 1865 1866 /* On failure, drop our mapping reference. */ 1867 VM_OBJECT_WLOCK(obj); 1868 shmfd->shm_kmappings--; 1869 VM_OBJECT_WUNLOCK(obj); 1870 1871 return (vm_mmap_to_errno(rv)); 1872 } 1873 1874 /* 1875 * We require the caller to unmap the entire entry. This allows us to 1876 * safely decrement shm_kmappings when a mapping is removed. 1877 */ 1878 int 1879 shm_unmap(struct file *fp, void *mem, size_t size) 1880 { 1881 struct shmfd *shmfd; 1882 vm_map_entry_t entry; 1883 vm_offset_t kva, ofs; 1884 vm_object_t obj; 1885 vm_pindex_t pindex; 1886 vm_prot_t prot; 1887 boolean_t wired; 1888 vm_map_t map; 1889 int rv; 1890 1891 if (fp->f_type != DTYPE_SHM) 1892 return (EINVAL); 1893 shmfd = fp->f_data; 1894 kva = (vm_offset_t)mem; 1895 ofs = kva & PAGE_MASK; 1896 kva = trunc_page(kva); 1897 size = round_page(size + ofs); 1898 map = kernel_map; 1899 rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry, 1900 &obj, &pindex, &prot, &wired); 1901 if (rv != KERN_SUCCESS) 1902 return (EINVAL); 1903 if (entry->start != kva || entry->end != kva + size) { 1904 vm_map_lookup_done(map, entry); 1905 return (EINVAL); 1906 } 1907 vm_map_lookup_done(map, entry); 1908 if (obj != shmfd->shm_object) 1909 return (EINVAL); 1910 vm_map_remove(map, kva, kva + size); 1911 VM_OBJECT_WLOCK(obj); 1912 KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped")); 1913 shmfd->shm_kmappings--; 1914 VM_OBJECT_WUNLOCK(obj); 1915 return (0); 1916 } 1917 1918 static int 1919 shm_fill_kinfo_locked(struct shmfd *shmfd, struct kinfo_file *kif, bool list) 1920 { 1921 const char *path, *pr_path; 1922 size_t pr_pathlen; 1923 bool visible; 1924 1925 sx_assert(&shm_dict_lock, SA_LOCKED); 1926 kif->kf_type = KF_TYPE_SHM; 1927 kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode; 1928 kif->kf_un.kf_file.kf_file_size = shmfd->shm_size; 1929 if (shmfd->shm_path != NULL) { 1930 path = shmfd->shm_path; 1931 pr_path = curthread->td_ucred->cr_prison->pr_path; 1932 if (strcmp(pr_path, "/") != 0) { 1933 /* Return the jail-rooted pathname. */ 1934 pr_pathlen = strlen(pr_path); 1935 visible = strncmp(path, pr_path, pr_pathlen) == 0 && 1936 path[pr_pathlen] == '/'; 1937 if (list && !visible) 1938 return (EPERM); 1939 if (visible) 1940 path += pr_pathlen; 1941 } 1942 strlcpy(kif->kf_path, path, sizeof(kif->kf_path)); 1943 } 1944 return (0); 1945 } 1946 1947 static int 1948 shm_fill_kinfo(struct file *fp, struct kinfo_file *kif, 1949 struct filedesc *fdp __unused) 1950 { 1951 int res; 1952 1953 sx_slock(&shm_dict_lock); 1954 res = shm_fill_kinfo_locked(fp->f_data, kif, false); 1955 sx_sunlock(&shm_dict_lock); 1956 return (res); 1957 } 1958 1959 static int 1960 shm_add_seals(struct file *fp, int seals) 1961 { 1962 struct shmfd *shmfd; 1963 void *rl_cookie; 1964 vm_ooffset_t writemappings; 1965 int error, nseals; 1966 1967 error = 0; 1968 shmfd = fp->f_data; 1969 rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX); 1970 1971 /* Even already-set seals should result in EPERM. */ 1972 if ((shmfd->shm_seals & F_SEAL_SEAL) != 0) { 1973 error = EPERM; 1974 goto out; 1975 } 1976 nseals = seals & ~shmfd->shm_seals; 1977 if ((nseals & F_SEAL_WRITE) != 0) { 1978 if (shm_largepage(shmfd)) { 1979 error = ENOTSUP; 1980 goto out; 1981 } 1982 1983 /* 1984 * The rangelock above prevents writable mappings from being 1985 * added after we've started applying seals. The RLOCK here 1986 * is to avoid torn reads on ILP32 arches as unmapping/reducing 1987 * writemappings will be done without a rangelock. 1988 */ 1989 VM_OBJECT_RLOCK(shmfd->shm_object); 1990 writemappings = shmfd->shm_object->un_pager.swp.writemappings; 1991 VM_OBJECT_RUNLOCK(shmfd->shm_object); 1992 /* kmappings are also writable */ 1993 if (writemappings > 0) { 1994 error = EBUSY; 1995 goto out; 1996 } 1997 } 1998 shmfd->shm_seals |= nseals; 1999 out: 2000 shm_rangelock_unlock(shmfd, rl_cookie); 2001 return (error); 2002 } 2003 2004 static int 2005 shm_get_seals(struct file *fp, int *seals) 2006 { 2007 struct shmfd *shmfd; 2008 2009 shmfd = fp->f_data; 2010 *seals = shmfd->shm_seals; 2011 return (0); 2012 } 2013 2014 static int 2015 shm_deallocate(struct shmfd *shmfd, off_t *offset, off_t *length, int flags) 2016 { 2017 vm_object_t object; 2018 vm_pindex_t pistart, pi, piend; 2019 vm_ooffset_t off, len; 2020 int startofs, endofs, end; 2021 int error; 2022 2023 off = *offset; 2024 len = *length; 2025 KASSERT(off + len <= (vm_ooffset_t)OFF_MAX, ("off + len overflows")); 2026 if (off + len > shmfd->shm_size) 2027 len = shmfd->shm_size - off; 2028 object = shmfd->shm_object; 2029 startofs = off & PAGE_MASK; 2030 endofs = (off + len) & PAGE_MASK; 2031 pistart = OFF_TO_IDX(off); 2032 piend = OFF_TO_IDX(off + len); 2033 pi = OFF_TO_IDX(off + PAGE_MASK); 2034 error = 0; 2035 2036 /* Handle the case when offset is on or beyond shm size. */ 2037 if ((off_t)len <= 0) { 2038 *length = 0; 2039 return (0); 2040 } 2041 2042 VM_OBJECT_WLOCK(object); 2043 2044 if (startofs != 0) { 2045 end = pistart != piend ? PAGE_SIZE : endofs; 2046 error = shm_partial_page_invalidate(object, pistart, startofs, 2047 end); 2048 if (error) 2049 goto out; 2050 off += end - startofs; 2051 len -= end - startofs; 2052 } 2053 2054 if (pi < piend) { 2055 vm_object_page_remove(object, pi, piend, 0); 2056 off += IDX_TO_OFF(piend - pi); 2057 len -= IDX_TO_OFF(piend - pi); 2058 } 2059 2060 if (endofs != 0 && pistart != piend) { 2061 error = shm_partial_page_invalidate(object, piend, 0, endofs); 2062 if (error) 2063 goto out; 2064 off += endofs; 2065 len -= endofs; 2066 } 2067 2068 out: 2069 VM_OBJECT_WUNLOCK(shmfd->shm_object); 2070 *offset = off; 2071 *length = len; 2072 return (error); 2073 } 2074 2075 static int 2076 shm_fspacectl(struct file *fp, int cmd, off_t *offset, off_t *length, int flags, 2077 struct ucred *active_cred, struct thread *td) 2078 { 2079 void *rl_cookie; 2080 struct shmfd *shmfd; 2081 off_t off, len; 2082 int error; 2083 2084 KASSERT(cmd == SPACECTL_DEALLOC, ("shm_fspacectl: Invalid cmd")); 2085 KASSERT((flags & ~SPACECTL_F_SUPPORTED) == 0, 2086 ("shm_fspacectl: non-zero flags")); 2087 KASSERT(*offset >= 0 && *length > 0 && *length <= OFF_MAX - *offset, 2088 ("shm_fspacectl: offset/length overflow or underflow")); 2089 error = EINVAL; 2090 shmfd = fp->f_data; 2091 off = *offset; 2092 len = *length; 2093 2094 rl_cookie = shm_rangelock_wlock(shmfd, off, off + len); 2095 switch (cmd) { 2096 case SPACECTL_DEALLOC: 2097 if ((shmfd->shm_seals & F_SEAL_WRITE) != 0) { 2098 error = EPERM; 2099 break; 2100 } 2101 error = shm_deallocate(shmfd, &off, &len, flags); 2102 *offset = off; 2103 *length = len; 2104 break; 2105 default: 2106 __assert_unreachable(); 2107 } 2108 shm_rangelock_unlock(shmfd, rl_cookie); 2109 return (error); 2110 } 2111 2112 2113 static int 2114 shm_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td) 2115 { 2116 void *rl_cookie; 2117 struct shmfd *shmfd; 2118 size_t size; 2119 int error; 2120 2121 /* This assumes that the caller already checked for overflow. */ 2122 error = 0; 2123 shmfd = fp->f_data; 2124 size = offset + len; 2125 2126 /* 2127 * Just grab the rangelock for the range that we may be attempting to 2128 * grow, rather than blocking read/write for regions we won't be 2129 * touching while this (potential) resize is in progress. Other 2130 * attempts to resize the shmfd will have to take a write lock from 0 to 2131 * OFF_MAX, so this being potentially beyond the current usable range of 2132 * the shmfd is not necessarily a concern. If other mechanisms are 2133 * added to grow a shmfd, this may need to be re-evaluated. 2134 */ 2135 rl_cookie = shm_rangelock_wlock(shmfd, offset, size); 2136 if (size > shmfd->shm_size) 2137 error = shm_dotruncate_cookie(shmfd, size, rl_cookie); 2138 shm_rangelock_unlock(shmfd, rl_cookie); 2139 /* Translate to posix_fallocate(2) return value as needed. */ 2140 if (error == ENOMEM) 2141 error = ENOSPC; 2142 return (error); 2143 } 2144 2145 static int 2146 sysctl_posix_shm_list(SYSCTL_HANDLER_ARGS) 2147 { 2148 struct shm_mapping *shmm; 2149 struct sbuf sb; 2150 struct kinfo_file kif; 2151 u_long i; 2152 int error, error2; 2153 2154 sbuf_new_for_sysctl(&sb, NULL, sizeof(struct kinfo_file) * 5, req); 2155 sbuf_clear_flags(&sb, SBUF_INCLUDENUL); 2156 error = 0; 2157 sx_slock(&shm_dict_lock); 2158 for (i = 0; i < shm_hash + 1; i++) { 2159 LIST_FOREACH(shmm, &shm_dictionary[i], sm_link) { 2160 error = shm_fill_kinfo_locked(shmm->sm_shmfd, 2161 &kif, true); 2162 if (error == EPERM) { 2163 error = 0; 2164 continue; 2165 } 2166 if (error != 0) 2167 break; 2168 pack_kinfo(&kif); 2169 error = sbuf_bcat(&sb, &kif, kif.kf_structsize) == 0 ? 2170 0 : ENOMEM; 2171 if (error != 0) 2172 break; 2173 } 2174 } 2175 sx_sunlock(&shm_dict_lock); 2176 error2 = sbuf_finish(&sb); 2177 sbuf_delete(&sb); 2178 return (error != 0 ? error : error2); 2179 } 2180 2181 SYSCTL_PROC(_kern_ipc, OID_AUTO, posix_shm_list, 2182 CTLFLAG_RD | CTLFLAG_PRISON | CTLFLAG_MPSAFE | CTLTYPE_OPAQUE, 2183 NULL, 0, sysctl_posix_shm_list, "", 2184 "POSIX SHM list"); 2185 2186 int 2187 kern_shm_open(struct thread *td, const char *path, int flags, mode_t mode, 2188 struct filecaps *caps) 2189 { 2190 2191 return (kern_shm_open2(td, path, flags, mode, 0, caps, NULL)); 2192 } 2193 2194 /* 2195 * This version of the shm_open() interface leaves CLOEXEC behavior up to the 2196 * caller, and libc will enforce it for the traditional shm_open() call. This 2197 * allows other consumers, like memfd_create(), to opt-in for CLOEXEC. This 2198 * interface also includes a 'name' argument that is currently unused, but could 2199 * potentially be exported later via some interface for debugging purposes. 2200 * From the kernel's perspective, it is optional. Individual consumers like 2201 * memfd_create() may require it in order to be compatible with other systems 2202 * implementing the same function. 2203 */ 2204 int 2205 sys_shm_open2(struct thread *td, struct shm_open2_args *uap) 2206 { 2207 2208 return (kern_shm_open2(td, uap->path, uap->flags, uap->mode, 2209 uap->shmflags, NULL, uap->name)); 2210 } 2211 2212 int 2213 shm_get_path(struct vm_object *obj, char *path, size_t sz) 2214 { 2215 struct shmfd *shmfd; 2216 int error; 2217 2218 error = 0; 2219 shmfd = NULL; 2220 sx_slock(&shm_dict_lock); 2221 VM_OBJECT_RLOCK(obj); 2222 if ((obj->flags & OBJ_POSIXSHM) == 0) { 2223 error = EINVAL; 2224 } else { 2225 if (obj->type == shmfd_pager_type) 2226 shmfd = obj->un_pager.swp.swp_priv; 2227 else if (obj->type == OBJT_PHYS) 2228 shmfd = obj->un_pager.phys.phys_priv; 2229 if (shmfd == NULL) { 2230 error = ENXIO; 2231 } else { 2232 strlcpy(path, shmfd->shm_path == NULL ? "anon" : 2233 shmfd->shm_path, sz); 2234 } 2235 } 2236 if (error != 0) 2237 path[0] = '\0'; 2238 VM_OBJECT_RUNLOCK(obj); 2239 sx_sunlock(&shm_dict_lock); 2240 return (error); 2241 } 2242