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 const 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 if (largepage) { 943 obj = phys_pager_allocate(NULL, &shm_largepage_phys_ops, 944 NULL, 0, VM_PROT_DEFAULT, 0, ucred); 945 } else { 946 obj = vm_pager_allocate(shmfd_pager_type, NULL, 0, 947 VM_PROT_DEFAULT, 0, ucred); 948 } 949 if (obj == NULL) { 950 /* 951 * swap reservation limits can cause object allocation 952 * to fail. 953 */ 954 return (NULL); 955 } 956 957 shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO); 958 shmfd->shm_uid = ucred->cr_uid; 959 shmfd->shm_gid = ucred->cr_gid; 960 shmfd->shm_mode = mode; 961 if (largepage) { 962 obj->un_pager.phys.phys_priv = shmfd; 963 shmfd->shm_lp_alloc_policy = SHM_LARGEPAGE_ALLOC_DEFAULT; 964 } else { 965 obj->un_pager.swp.swp_priv = shmfd; 966 } 967 968 VM_OBJECT_WLOCK(obj); 969 vm_object_set_flag(obj, OBJ_POSIXSHM); 970 VM_OBJECT_WUNLOCK(obj); 971 shmfd->shm_object = obj; 972 vfs_timestamp(&shmfd->shm_birthtime); 973 shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime = 974 shmfd->shm_birthtime; 975 shmfd->shm_ino = alloc_unr64(&shm_ino_unr); 976 refcount_init(&shmfd->shm_refs, 1); 977 mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF); 978 rangelock_init(&shmfd->shm_rl); 979 #ifdef MAC 980 mac_posixshm_init(shmfd); 981 mac_posixshm_create(ucred, shmfd); 982 #endif 983 984 return (shmfd); 985 } 986 987 struct shmfd * 988 shm_hold(struct shmfd *shmfd) 989 { 990 991 refcount_acquire(&shmfd->shm_refs); 992 return (shmfd); 993 } 994 995 void 996 shm_drop(struct shmfd *shmfd) 997 { 998 vm_object_t obj; 999 1000 if (refcount_release(&shmfd->shm_refs)) { 1001 #ifdef MAC 1002 mac_posixshm_destroy(shmfd); 1003 #endif 1004 rangelock_destroy(&shmfd->shm_rl); 1005 mtx_destroy(&shmfd->shm_mtx); 1006 obj = shmfd->shm_object; 1007 VM_OBJECT_WLOCK(obj); 1008 if (shm_largepage(shmfd)) 1009 obj->un_pager.phys.phys_priv = NULL; 1010 else 1011 obj->un_pager.swp.swp_priv = NULL; 1012 VM_OBJECT_WUNLOCK(obj); 1013 vm_object_deallocate(obj); 1014 free(shmfd, M_SHMFD); 1015 } 1016 } 1017 1018 /* 1019 * Determine if the credentials have sufficient permissions for a 1020 * specified combination of FREAD and FWRITE. 1021 */ 1022 int 1023 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags) 1024 { 1025 accmode_t accmode; 1026 int error; 1027 1028 accmode = 0; 1029 if (flags & FREAD) 1030 accmode |= VREAD; 1031 if (flags & FWRITE) 1032 accmode |= VWRITE; 1033 mtx_lock(&shm_timestamp_lock); 1034 error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid, 1035 accmode, ucred); 1036 mtx_unlock(&shm_timestamp_lock); 1037 return (error); 1038 } 1039 1040 static void 1041 shm_init(void *arg) 1042 { 1043 char name[32]; 1044 int i; 1045 1046 mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF); 1047 sx_init(&shm_dict_lock, "shm dictionary"); 1048 shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash); 1049 new_unrhdr64(&shm_ino_unr, 1); 1050 shm_dev_ino = devfs_alloc_cdp_inode(); 1051 KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized")); 1052 shmfd_pager_type = vm_pager_alloc_dyn_type(&shm_swap_pager_ops, 1053 OBJT_SWAP); 1054 MPASS(shmfd_pager_type != -1); 1055 1056 for (i = 1; i < MAXPAGESIZES; i++) { 1057 if (pagesizes[i] == 0) 1058 break; 1059 #define M (1024 * 1024) 1060 #define G (1024 * M) 1061 if (pagesizes[i] >= G) 1062 snprintf(name, sizeof(name), "%luG", pagesizes[i] / G); 1063 else if (pagesizes[i] >= M) 1064 snprintf(name, sizeof(name), "%luM", pagesizes[i] / M); 1065 else 1066 snprintf(name, sizeof(name), "%lu", pagesizes[i]); 1067 #undef G 1068 #undef M 1069 SYSCTL_ADD_ULONG(NULL, SYSCTL_STATIC_CHILDREN(_vm_largepages), 1070 OID_AUTO, name, CTLFLAG_RD, &count_largepages[i], 1071 "number of non-transient largepages allocated"); 1072 } 1073 } 1074 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL); 1075 1076 /* 1077 * Remove all shared memory objects that belong to a prison. 1078 */ 1079 void 1080 shm_remove_prison(struct prison *pr) 1081 { 1082 struct shm_mapping *shmm, *tshmm; 1083 u_long i; 1084 1085 sx_xlock(&shm_dict_lock); 1086 for (i = 0; i < shm_hash + 1; i++) { 1087 LIST_FOREACH_SAFE(shmm, &shm_dictionary[i], sm_link, tshmm) { 1088 if (shmm->sm_shmfd->shm_object->cred && 1089 shmm->sm_shmfd->shm_object->cred->cr_prison == pr) 1090 shm_doremove(shmm); 1091 } 1092 } 1093 sx_xunlock(&shm_dict_lock); 1094 } 1095 1096 /* 1097 * Dictionary management. We maintain an in-kernel dictionary to map 1098 * paths to shmfd objects. We use the FNV hash on the path to store 1099 * the mappings in a hash table. 1100 */ 1101 static struct shmfd * 1102 shm_lookup(char *path, Fnv32_t fnv) 1103 { 1104 struct shm_mapping *map; 1105 1106 LIST_FOREACH(map, SHM_HASH(fnv), sm_link) { 1107 if (map->sm_fnv != fnv) 1108 continue; 1109 if (strcmp(map->sm_path, path) == 0) 1110 return (map->sm_shmfd); 1111 } 1112 1113 return (NULL); 1114 } 1115 1116 static void 1117 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd) 1118 { 1119 struct shm_mapping *map; 1120 1121 map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK); 1122 map->sm_path = path; 1123 map->sm_fnv = fnv; 1124 map->sm_shmfd = shm_hold(shmfd); 1125 shmfd->shm_path = path; 1126 LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link); 1127 } 1128 1129 static int 1130 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred) 1131 { 1132 struct shm_mapping *map; 1133 int error; 1134 1135 LIST_FOREACH(map, SHM_HASH(fnv), sm_link) { 1136 if (map->sm_fnv != fnv) 1137 continue; 1138 if (strcmp(map->sm_path, path) == 0) { 1139 #ifdef MAC 1140 error = mac_posixshm_check_unlink(ucred, map->sm_shmfd); 1141 if (error) 1142 return (error); 1143 #endif 1144 error = shm_access(map->sm_shmfd, ucred, 1145 FREAD | FWRITE); 1146 if (error) 1147 return (error); 1148 shm_doremove(map); 1149 return (0); 1150 } 1151 } 1152 1153 return (ENOENT); 1154 } 1155 1156 static void 1157 shm_doremove(struct shm_mapping *map) 1158 { 1159 map->sm_shmfd->shm_path = NULL; 1160 LIST_REMOVE(map, sm_link); 1161 shm_drop(map->sm_shmfd); 1162 free(map->sm_path, M_SHMFD); 1163 free(map, M_SHMFD); 1164 } 1165 1166 int 1167 kern_shm_open2(struct thread *td, const char *userpath, int flags, mode_t mode, 1168 int shmflags, struct filecaps *fcaps, const char *name __unused) 1169 { 1170 struct pwddesc *pdp; 1171 struct shmfd *shmfd; 1172 struct file *fp; 1173 char *path; 1174 void *rl_cookie; 1175 Fnv32_t fnv; 1176 mode_t cmode; 1177 int error, fd, initial_seals; 1178 bool largepage; 1179 1180 if ((shmflags & ~(SHM_ALLOW_SEALING | SHM_GROW_ON_WRITE | 1181 SHM_LARGEPAGE)) != 0) 1182 return (EINVAL); 1183 1184 initial_seals = F_SEAL_SEAL; 1185 if ((shmflags & SHM_ALLOW_SEALING) != 0) 1186 initial_seals &= ~F_SEAL_SEAL; 1187 1188 AUDIT_ARG_FFLAGS(flags); 1189 AUDIT_ARG_MODE(mode); 1190 1191 if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR) 1192 return (EINVAL); 1193 1194 if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0) 1195 return (EINVAL); 1196 1197 largepage = (shmflags & SHM_LARGEPAGE) != 0; 1198 if (largepage && !PMAP_HAS_LARGEPAGES) 1199 return (ENOTTY); 1200 1201 /* 1202 * Currently only F_SEAL_SEAL may be set when creating or opening shmfd. 1203 * If the decision is made later to allow additional seals, care must be 1204 * taken below to ensure that the seals are properly set if the shmfd 1205 * already existed -- this currently assumes that only F_SEAL_SEAL can 1206 * be set and doesn't take further precautions to ensure the validity of 1207 * the seals being added with respect to current mappings. 1208 */ 1209 if ((initial_seals & ~F_SEAL_SEAL) != 0) 1210 return (EINVAL); 1211 1212 if (userpath != SHM_ANON) { 1213 error = shm_copyin_path(td, userpath, &path); 1214 if (error != 0) 1215 return (error); 1216 1217 #ifdef CAPABILITY_MODE 1218 /* 1219 * shm_open(2) is only allowed for anonymous objects. 1220 */ 1221 if (CAP_TRACING(td)) 1222 ktrcapfail(CAPFAIL_NAMEI, path); 1223 if (IN_CAPABILITY_MODE(td)) { 1224 error = ECAPMODE; 1225 goto outnofp; 1226 } 1227 #endif 1228 1229 AUDIT_ARG_UPATH1_CANON(path); 1230 } else { 1231 path = NULL; 1232 } 1233 1234 pdp = td->td_proc->p_pd; 1235 cmode = (mode & ~pdp->pd_cmask) & ACCESSPERMS; 1236 1237 /* 1238 * shm_open(2) created shm should always have O_CLOEXEC set, as mandated 1239 * by POSIX. We allow it to be unset here so that an in-kernel 1240 * interface may be written as a thin layer around shm, optionally not 1241 * setting CLOEXEC. For shm_open(2), O_CLOEXEC is set unconditionally 1242 * in sys_shm_open() to keep this implementation compliant. 1243 */ 1244 error = falloc_caps(td, &fp, &fd, flags & O_CLOEXEC, fcaps); 1245 if (error != 0) 1246 goto outnofp; 1247 1248 /* A SHM_ANON path pointer creates an anonymous object. */ 1249 if (userpath == SHM_ANON) { 1250 /* A read-only anonymous object is pointless. */ 1251 if ((flags & O_ACCMODE) == O_RDONLY) { 1252 error = EINVAL; 1253 goto out; 1254 } 1255 shmfd = shm_alloc(td->td_ucred, cmode, largepage); 1256 if (shmfd == NULL) { 1257 error = ENOMEM; 1258 goto out; 1259 } 1260 shmfd->shm_seals = initial_seals; 1261 shmfd->shm_flags = shmflags; 1262 } else { 1263 fnv = fnv_32_str(path, FNV1_32_INIT); 1264 sx_xlock(&shm_dict_lock); 1265 shmfd = shm_lookup(path, fnv); 1266 if (shmfd == NULL) { 1267 /* Object does not yet exist, create it if requested. */ 1268 if (flags & O_CREAT) { 1269 #ifdef MAC 1270 error = mac_posixshm_check_create(td->td_ucred, 1271 path); 1272 if (error == 0) { 1273 #endif 1274 shmfd = shm_alloc(td->td_ucred, cmode, 1275 largepage); 1276 if (shmfd == NULL) { 1277 error = ENOMEM; 1278 } else { 1279 shmfd->shm_seals = 1280 initial_seals; 1281 shmfd->shm_flags = shmflags; 1282 shm_insert(path, fnv, shmfd); 1283 path = NULL; 1284 } 1285 #ifdef MAC 1286 } 1287 #endif 1288 } else { 1289 error = ENOENT; 1290 } 1291 } else { 1292 /* 1293 * Object already exists, obtain a new reference if 1294 * requested and permitted. 1295 */ 1296 rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX); 1297 1298 /* 1299 * kern_shm_open() likely shouldn't ever error out on 1300 * trying to set a seal that already exists, unlike 1301 * F_ADD_SEALS. This would break terribly as 1302 * shm_open(2) actually sets F_SEAL_SEAL to maintain 1303 * historical behavior where the underlying file could 1304 * not be sealed. 1305 */ 1306 initial_seals &= ~shmfd->shm_seals; 1307 1308 /* 1309 * initial_seals can't set additional seals if we've 1310 * already been set F_SEAL_SEAL. If F_SEAL_SEAL is set, 1311 * then we've already removed that one from 1312 * initial_seals. This is currently redundant as we 1313 * only allow setting F_SEAL_SEAL at creation time, but 1314 * it's cheap to check and decreases the effort required 1315 * to allow additional seals. 1316 */ 1317 if ((shmfd->shm_seals & F_SEAL_SEAL) != 0 && 1318 initial_seals != 0) 1319 error = EPERM; 1320 else if ((flags & (O_CREAT | O_EXCL)) == 1321 (O_CREAT | O_EXCL)) 1322 error = EEXIST; 1323 else if (shmflags != 0 && shmflags != shmfd->shm_flags) 1324 error = EINVAL; 1325 else { 1326 #ifdef MAC 1327 error = mac_posixshm_check_open(td->td_ucred, 1328 shmfd, FFLAGS(flags & O_ACCMODE)); 1329 if (error == 0) 1330 #endif 1331 error = shm_access(shmfd, td->td_ucred, 1332 FFLAGS(flags & O_ACCMODE)); 1333 } 1334 1335 /* 1336 * Truncate the file back to zero length if 1337 * O_TRUNC was specified and the object was 1338 * opened with read/write. 1339 */ 1340 if (error == 0 && 1341 (flags & (O_ACCMODE | O_TRUNC)) == 1342 (O_RDWR | O_TRUNC)) { 1343 VM_OBJECT_WLOCK(shmfd->shm_object); 1344 #ifdef MAC 1345 error = mac_posixshm_check_truncate( 1346 td->td_ucred, fp->f_cred, shmfd); 1347 if (error == 0) 1348 #endif 1349 error = shm_dotruncate_locked(shmfd, 0, 1350 rl_cookie); 1351 VM_OBJECT_WUNLOCK(shmfd->shm_object); 1352 } 1353 if (error == 0) { 1354 /* 1355 * Currently we only allow F_SEAL_SEAL to be 1356 * set initially. As noted above, this would 1357 * need to be reworked should that change. 1358 */ 1359 shmfd->shm_seals |= initial_seals; 1360 shm_hold(shmfd); 1361 } 1362 shm_rangelock_unlock(shmfd, rl_cookie); 1363 } 1364 sx_xunlock(&shm_dict_lock); 1365 1366 if (error != 0) 1367 goto out; 1368 } 1369 1370 finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops); 1371 1372 td->td_retval[0] = fd; 1373 fdrop(fp, td); 1374 free(path, M_SHMFD); 1375 1376 return (0); 1377 1378 out: 1379 fdclose(td, fp, fd); 1380 fdrop(fp, td); 1381 outnofp: 1382 free(path, M_SHMFD); 1383 1384 return (error); 1385 } 1386 1387 /* System calls. */ 1388 #ifdef COMPAT_FREEBSD12 1389 int 1390 freebsd12_shm_open(struct thread *td, struct freebsd12_shm_open_args *uap) 1391 { 1392 1393 return (kern_shm_open(td, uap->path, uap->flags | O_CLOEXEC, 1394 uap->mode, NULL)); 1395 } 1396 #endif 1397 1398 int 1399 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap) 1400 { 1401 char *path; 1402 Fnv32_t fnv; 1403 int error; 1404 1405 error = shm_copyin_path(td, uap->path, &path); 1406 if (error != 0) 1407 return (error); 1408 1409 AUDIT_ARG_UPATH1_CANON(path); 1410 fnv = fnv_32_str(path, FNV1_32_INIT); 1411 sx_xlock(&shm_dict_lock); 1412 error = shm_remove(path, fnv, td->td_ucred); 1413 sx_xunlock(&shm_dict_lock); 1414 free(path, M_SHMFD); 1415 1416 return (error); 1417 } 1418 1419 int 1420 sys_shm_rename(struct thread *td, struct shm_rename_args *uap) 1421 { 1422 char *path_from = NULL, *path_to = NULL; 1423 Fnv32_t fnv_from, fnv_to; 1424 struct shmfd *fd_from; 1425 struct shmfd *fd_to; 1426 int error; 1427 int flags; 1428 1429 flags = uap->flags; 1430 AUDIT_ARG_FFLAGS(flags); 1431 1432 /* 1433 * Make sure the user passed only valid flags. 1434 * If you add a new flag, please add a new term here. 1435 */ 1436 if ((flags & ~( 1437 SHM_RENAME_NOREPLACE | 1438 SHM_RENAME_EXCHANGE 1439 )) != 0) { 1440 error = EINVAL; 1441 goto out; 1442 } 1443 1444 /* 1445 * EXCHANGE and NOREPLACE don't quite make sense together. Let's 1446 * force the user to choose one or the other. 1447 */ 1448 if ((flags & SHM_RENAME_NOREPLACE) != 0 && 1449 (flags & SHM_RENAME_EXCHANGE) != 0) { 1450 error = EINVAL; 1451 goto out; 1452 } 1453 1454 /* Renaming to or from anonymous makes no sense */ 1455 if (uap->path_from == SHM_ANON || uap->path_to == SHM_ANON) { 1456 error = EINVAL; 1457 goto out; 1458 } 1459 1460 error = shm_copyin_path(td, uap->path_from, &path_from); 1461 if (error != 0) 1462 goto out; 1463 1464 error = shm_copyin_path(td, uap->path_to, &path_to); 1465 if (error != 0) 1466 goto out; 1467 1468 AUDIT_ARG_UPATH1_CANON(path_from); 1469 AUDIT_ARG_UPATH2_CANON(path_to); 1470 1471 /* Rename with from/to equal is a no-op */ 1472 if (strcmp(path_from, path_to) == 0) 1473 goto out; 1474 1475 fnv_from = fnv_32_str(path_from, FNV1_32_INIT); 1476 fnv_to = fnv_32_str(path_to, FNV1_32_INIT); 1477 1478 sx_xlock(&shm_dict_lock); 1479 1480 fd_from = shm_lookup(path_from, fnv_from); 1481 if (fd_from == NULL) { 1482 error = ENOENT; 1483 goto out_locked; 1484 } 1485 1486 fd_to = shm_lookup(path_to, fnv_to); 1487 if ((flags & SHM_RENAME_NOREPLACE) != 0 && fd_to != NULL) { 1488 error = EEXIST; 1489 goto out_locked; 1490 } 1491 1492 /* 1493 * Unconditionally prevents shm_remove from invalidating the 'from' 1494 * shm's state. 1495 */ 1496 shm_hold(fd_from); 1497 error = shm_remove(path_from, fnv_from, td->td_ucred); 1498 1499 /* 1500 * One of my assumptions failed if ENOENT (e.g. locking didn't 1501 * protect us) 1502 */ 1503 KASSERT(error != ENOENT, ("Our shm disappeared during shm_rename: %s", 1504 path_from)); 1505 if (error != 0) { 1506 shm_drop(fd_from); 1507 goto out_locked; 1508 } 1509 1510 /* 1511 * If we are exchanging, we need to ensure the shm_remove below 1512 * doesn't invalidate the dest shm's state. 1513 */ 1514 if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) 1515 shm_hold(fd_to); 1516 1517 /* 1518 * NOTE: if path_to is not already in the hash, c'est la vie; 1519 * it simply means we have nothing already at path_to to unlink. 1520 * That is the ENOENT case. 1521 * 1522 * If we somehow don't have access to unlink this guy, but 1523 * did for the shm at path_from, then relink the shm to path_from 1524 * and abort with EACCES. 1525 * 1526 * All other errors: that is weird; let's relink and abort the 1527 * operation. 1528 */ 1529 error = shm_remove(path_to, fnv_to, td->td_ucred); 1530 if (error != 0 && error != ENOENT) { 1531 shm_insert(path_from, fnv_from, fd_from); 1532 shm_drop(fd_from); 1533 /* Don't free path_from now, since the hash references it */ 1534 path_from = NULL; 1535 goto out_locked; 1536 } 1537 1538 error = 0; 1539 1540 shm_insert(path_to, fnv_to, fd_from); 1541 1542 /* Don't free path_to now, since the hash references it */ 1543 path_to = NULL; 1544 1545 /* We kept a ref when we removed, and incremented again in insert */ 1546 shm_drop(fd_from); 1547 KASSERT(fd_from->shm_refs > 0, ("Expected >0 refs; got: %d\n", 1548 fd_from->shm_refs)); 1549 1550 if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) { 1551 shm_insert(path_from, fnv_from, fd_to); 1552 path_from = NULL; 1553 shm_drop(fd_to); 1554 KASSERT(fd_to->shm_refs > 0, ("Expected >0 refs; got: %d\n", 1555 fd_to->shm_refs)); 1556 } 1557 1558 out_locked: 1559 sx_xunlock(&shm_dict_lock); 1560 1561 out: 1562 free(path_from, M_SHMFD); 1563 free(path_to, M_SHMFD); 1564 return (error); 1565 } 1566 1567 static int 1568 shm_mmap_large(struct shmfd *shmfd, vm_map_t map, vm_offset_t *addr, 1569 vm_size_t size, vm_prot_t prot, vm_prot_t max_prot, int flags, 1570 vm_ooffset_t foff, struct thread *td) 1571 { 1572 struct vmspace *vms; 1573 vm_map_entry_t next_entry, prev_entry; 1574 vm_offset_t align, mask, maxaddr; 1575 int docow, error, rv, try; 1576 bool curmap; 1577 1578 if (shmfd->shm_lp_psind == 0) 1579 return (EINVAL); 1580 1581 /* MAP_PRIVATE is disabled */ 1582 if ((flags & ~(MAP_SHARED | MAP_FIXED | MAP_EXCL | 1583 MAP_NOCORE | MAP_32BIT | MAP_ALIGNMENT_MASK)) != 0) 1584 return (EINVAL); 1585 1586 vms = td->td_proc->p_vmspace; 1587 curmap = map == &vms->vm_map; 1588 if (curmap) { 1589 error = kern_mmap_racct_check(td, map, size); 1590 if (error != 0) 1591 return (error); 1592 } 1593 1594 docow = shmfd->shm_lp_psind << MAP_SPLIT_BOUNDARY_SHIFT; 1595 docow |= MAP_INHERIT_SHARE; 1596 if ((flags & MAP_NOCORE) != 0) 1597 docow |= MAP_DISABLE_COREDUMP; 1598 1599 mask = pagesizes[shmfd->shm_lp_psind] - 1; 1600 if ((foff & mask) != 0) 1601 return (EINVAL); 1602 maxaddr = vm_map_max(map); 1603 if ((flags & MAP_32BIT) != 0 && maxaddr > MAP_32BIT_MAX_ADDR) 1604 maxaddr = MAP_32BIT_MAX_ADDR; 1605 if (size == 0 || (size & mask) != 0 || 1606 (*addr != 0 && ((*addr & mask) != 0 || 1607 *addr + size < *addr || *addr + size > maxaddr))) 1608 return (EINVAL); 1609 1610 align = flags & MAP_ALIGNMENT_MASK; 1611 if (align == 0) { 1612 align = pagesizes[shmfd->shm_lp_psind]; 1613 } else if (align == MAP_ALIGNED_SUPER) { 1614 /* 1615 * MAP_ALIGNED_SUPER is only supported on superpage sizes, 1616 * i.e., [1, VM_NRESERVLEVEL]. shmfd->shm_lp_psind < 1 is 1617 * handled above. 1618 */ 1619 if ( 1620 #if VM_NRESERVLEVEL > 0 1621 shmfd->shm_lp_psind > VM_NRESERVLEVEL 1622 #else 1623 shmfd->shm_lp_psind > 1 1624 #endif 1625 ) 1626 return (EINVAL); 1627 align = pagesizes[shmfd->shm_lp_psind]; 1628 } else { 1629 align >>= MAP_ALIGNMENT_SHIFT; 1630 align = 1ULL << align; 1631 /* Also handles overflow. */ 1632 if (align < pagesizes[shmfd->shm_lp_psind]) 1633 return (EINVAL); 1634 } 1635 1636 vm_map_lock(map); 1637 if ((flags & MAP_FIXED) == 0) { 1638 try = 1; 1639 if (curmap && (*addr == 0 || 1640 (*addr >= round_page((vm_offset_t)vms->vm_taddr) && 1641 *addr < round_page((vm_offset_t)vms->vm_daddr + 1642 lim_max(td, RLIMIT_DATA))))) { 1643 *addr = roundup2((vm_offset_t)vms->vm_daddr + 1644 lim_max(td, RLIMIT_DATA), 1645 pagesizes[shmfd->shm_lp_psind]); 1646 } 1647 again: 1648 rv = vm_map_find_aligned(map, addr, size, maxaddr, align); 1649 if (rv != KERN_SUCCESS) { 1650 if (try == 1) { 1651 try = 2; 1652 *addr = vm_map_min(map); 1653 if ((*addr & mask) != 0) 1654 *addr = (*addr + mask) & mask; 1655 goto again; 1656 } 1657 goto fail1; 1658 } 1659 } else if ((flags & MAP_EXCL) == 0) { 1660 rv = vm_map_delete(map, *addr, *addr + size); 1661 if (rv != KERN_SUCCESS) 1662 goto fail1; 1663 } else { 1664 error = ENOSPC; 1665 if (vm_map_lookup_entry(map, *addr, &prev_entry)) 1666 goto fail; 1667 next_entry = vm_map_entry_succ(prev_entry); 1668 if (next_entry->start < *addr + size) 1669 goto fail; 1670 } 1671 1672 rv = vm_map_insert(map, shmfd->shm_object, foff, *addr, *addr + size, 1673 prot, max_prot, docow); 1674 fail1: 1675 error = vm_mmap_to_errno(rv); 1676 fail: 1677 vm_map_unlock(map); 1678 return (error); 1679 } 1680 1681 static int 1682 shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize, 1683 vm_prot_t prot, vm_prot_t max_maxprot, int flags, 1684 vm_ooffset_t foff, struct thread *td) 1685 { 1686 struct shmfd *shmfd; 1687 vm_prot_t maxprot; 1688 int error; 1689 bool writecnt; 1690 void *rl_cookie; 1691 1692 shmfd = fp->f_data; 1693 maxprot = VM_PROT_NONE; 1694 1695 rl_cookie = shm_rangelock_rlock(shmfd, 0, objsize); 1696 /* FREAD should always be set. */ 1697 if ((fp->f_flag & FREAD) != 0) 1698 maxprot |= VM_PROT_EXECUTE | VM_PROT_READ; 1699 1700 /* 1701 * If FWRITE's set, we can allow VM_PROT_WRITE unless it's a shared 1702 * mapping with a write seal applied. Private mappings are always 1703 * writeable. 1704 */ 1705 if ((flags & MAP_SHARED) == 0) { 1706 if ((max_maxprot & VM_PROT_WRITE) != 0) 1707 maxprot |= VM_PROT_WRITE; 1708 writecnt = false; 1709 } else { 1710 if ((fp->f_flag & FWRITE) != 0 && 1711 (shmfd->shm_seals & F_SEAL_WRITE) == 0) 1712 maxprot |= VM_PROT_WRITE; 1713 1714 /* 1715 * Any mappings from a writable descriptor may be upgraded to 1716 * VM_PROT_WRITE with mprotect(2), unless a write-seal was 1717 * applied between the open and subsequent mmap(2). We want to 1718 * reject application of a write seal as long as any such 1719 * mapping exists so that the seal cannot be trivially bypassed. 1720 */ 1721 writecnt = (maxprot & VM_PROT_WRITE) != 0; 1722 if (!writecnt && (prot & VM_PROT_WRITE) != 0) { 1723 error = EACCES; 1724 goto out; 1725 } 1726 } 1727 maxprot &= max_maxprot; 1728 1729 /* See comment in vn_mmap(). */ 1730 if ( 1731 #ifdef _LP64 1732 objsize > OFF_MAX || 1733 #endif 1734 foff > OFF_MAX - objsize) { 1735 error = EINVAL; 1736 goto out; 1737 } 1738 1739 #ifdef MAC 1740 error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags); 1741 if (error != 0) 1742 goto out; 1743 #endif 1744 1745 mtx_lock(&shm_timestamp_lock); 1746 vfs_timestamp(&shmfd->shm_atime); 1747 mtx_unlock(&shm_timestamp_lock); 1748 vm_object_reference(shmfd->shm_object); 1749 1750 if (shm_largepage(shmfd)) { 1751 writecnt = false; 1752 error = shm_mmap_large(shmfd, map, addr, objsize, prot, 1753 maxprot, flags, foff, td); 1754 } else { 1755 if (writecnt) { 1756 vm_pager_update_writecount(shmfd->shm_object, 0, 1757 objsize); 1758 } 1759 error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags, 1760 shmfd->shm_object, foff, writecnt, td); 1761 } 1762 if (error != 0) { 1763 if (writecnt) 1764 vm_pager_release_writecount(shmfd->shm_object, 0, 1765 objsize); 1766 vm_object_deallocate(shmfd->shm_object); 1767 } 1768 out: 1769 shm_rangelock_unlock(shmfd, rl_cookie); 1770 return (error); 1771 } 1772 1773 static int 1774 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, 1775 struct thread *td) 1776 { 1777 struct shmfd *shmfd; 1778 int error; 1779 1780 error = 0; 1781 shmfd = fp->f_data; 1782 mtx_lock(&shm_timestamp_lock); 1783 /* 1784 * SUSv4 says that x bits of permission need not be affected. 1785 * Be consistent with our shm_open there. 1786 */ 1787 #ifdef MAC 1788 error = mac_posixshm_check_setmode(active_cred, shmfd, mode); 1789 if (error != 0) 1790 goto out; 1791 #endif 1792 error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid, 1793 VADMIN, active_cred); 1794 if (error != 0) 1795 goto out; 1796 shmfd->shm_mode = mode & ACCESSPERMS; 1797 out: 1798 mtx_unlock(&shm_timestamp_lock); 1799 return (error); 1800 } 1801 1802 static int 1803 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred, 1804 struct thread *td) 1805 { 1806 struct shmfd *shmfd; 1807 int error; 1808 1809 error = 0; 1810 shmfd = fp->f_data; 1811 mtx_lock(&shm_timestamp_lock); 1812 #ifdef MAC 1813 error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid); 1814 if (error != 0) 1815 goto out; 1816 #endif 1817 if (uid == (uid_t)-1) 1818 uid = shmfd->shm_uid; 1819 if (gid == (gid_t)-1) 1820 gid = shmfd->shm_gid; 1821 if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) || 1822 (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) && 1823 (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN))) 1824 goto out; 1825 shmfd->shm_uid = uid; 1826 shmfd->shm_gid = gid; 1827 out: 1828 mtx_unlock(&shm_timestamp_lock); 1829 return (error); 1830 } 1831 1832 /* 1833 * Helper routines to allow the backing object of a shared memory file 1834 * descriptor to be mapped in the kernel. 1835 */ 1836 int 1837 shm_map(struct file *fp, size_t size, off_t offset, void **memp) 1838 { 1839 struct shmfd *shmfd; 1840 vm_offset_t kva, ofs; 1841 vm_object_t obj; 1842 int rv; 1843 1844 if (fp->f_type != DTYPE_SHM) 1845 return (EINVAL); 1846 shmfd = fp->f_data; 1847 obj = shmfd->shm_object; 1848 VM_OBJECT_WLOCK(obj); 1849 /* 1850 * XXXRW: This validation is probably insufficient, and subject to 1851 * sign errors. It should be fixed. 1852 */ 1853 if (offset >= shmfd->shm_size || 1854 offset + size > round_page(shmfd->shm_size)) { 1855 VM_OBJECT_WUNLOCK(obj); 1856 return (EINVAL); 1857 } 1858 1859 shmfd->shm_kmappings++; 1860 vm_object_reference_locked(obj); 1861 VM_OBJECT_WUNLOCK(obj); 1862 1863 /* Map the object into the kernel_map and wire it. */ 1864 kva = vm_map_min(kernel_map); 1865 ofs = offset & PAGE_MASK; 1866 offset = trunc_page(offset); 1867 size = round_page(size + ofs); 1868 rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0, 1869 VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE, 1870 VM_PROT_READ | VM_PROT_WRITE, 0); 1871 if (rv == KERN_SUCCESS) { 1872 rv = vm_map_wire(kernel_map, kva, kva + size, 1873 VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES); 1874 if (rv == KERN_SUCCESS) { 1875 *memp = (void *)(kva + ofs); 1876 return (0); 1877 } 1878 vm_map_remove(kernel_map, kva, kva + size); 1879 } else 1880 vm_object_deallocate(obj); 1881 1882 /* On failure, drop our mapping reference. */ 1883 VM_OBJECT_WLOCK(obj); 1884 shmfd->shm_kmappings--; 1885 VM_OBJECT_WUNLOCK(obj); 1886 1887 return (vm_mmap_to_errno(rv)); 1888 } 1889 1890 /* 1891 * We require the caller to unmap the entire entry. This allows us to 1892 * safely decrement shm_kmappings when a mapping is removed. 1893 */ 1894 int 1895 shm_unmap(struct file *fp, void *mem, size_t size) 1896 { 1897 struct shmfd *shmfd; 1898 vm_map_entry_t entry; 1899 vm_offset_t kva, ofs; 1900 vm_object_t obj; 1901 vm_pindex_t pindex; 1902 vm_prot_t prot; 1903 boolean_t wired; 1904 vm_map_t map; 1905 int rv; 1906 1907 if (fp->f_type != DTYPE_SHM) 1908 return (EINVAL); 1909 shmfd = fp->f_data; 1910 kva = (vm_offset_t)mem; 1911 ofs = kva & PAGE_MASK; 1912 kva = trunc_page(kva); 1913 size = round_page(size + ofs); 1914 map = kernel_map; 1915 rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry, 1916 &obj, &pindex, &prot, &wired); 1917 if (rv != KERN_SUCCESS) 1918 return (EINVAL); 1919 if (entry->start != kva || entry->end != kva + size) { 1920 vm_map_lookup_done(map, entry); 1921 return (EINVAL); 1922 } 1923 vm_map_lookup_done(map, entry); 1924 if (obj != shmfd->shm_object) 1925 return (EINVAL); 1926 vm_map_remove(map, kva, kva + size); 1927 VM_OBJECT_WLOCK(obj); 1928 KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped")); 1929 shmfd->shm_kmappings--; 1930 VM_OBJECT_WUNLOCK(obj); 1931 return (0); 1932 } 1933 1934 static int 1935 shm_fill_kinfo_locked(struct shmfd *shmfd, struct kinfo_file *kif, bool list) 1936 { 1937 const char *path, *pr_path; 1938 size_t pr_pathlen; 1939 bool visible; 1940 1941 sx_assert(&shm_dict_lock, SA_LOCKED); 1942 kif->kf_type = KF_TYPE_SHM; 1943 kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode; 1944 kif->kf_un.kf_file.kf_file_size = shmfd->shm_size; 1945 if (shmfd->shm_path != NULL) { 1946 path = shmfd->shm_path; 1947 pr_path = curthread->td_ucred->cr_prison->pr_path; 1948 if (strcmp(pr_path, "/") != 0) { 1949 /* Return the jail-rooted pathname. */ 1950 pr_pathlen = strlen(pr_path); 1951 visible = strncmp(path, pr_path, pr_pathlen) == 0 && 1952 path[pr_pathlen] == '/'; 1953 if (list && !visible) 1954 return (EPERM); 1955 if (visible) 1956 path += pr_pathlen; 1957 } 1958 strlcpy(kif->kf_path, path, sizeof(kif->kf_path)); 1959 } 1960 return (0); 1961 } 1962 1963 static int 1964 shm_fill_kinfo(struct file *fp, struct kinfo_file *kif, 1965 struct filedesc *fdp __unused) 1966 { 1967 int res; 1968 1969 sx_slock(&shm_dict_lock); 1970 res = shm_fill_kinfo_locked(fp->f_data, kif, false); 1971 sx_sunlock(&shm_dict_lock); 1972 return (res); 1973 } 1974 1975 static int 1976 shm_add_seals(struct file *fp, int seals) 1977 { 1978 struct shmfd *shmfd; 1979 void *rl_cookie; 1980 vm_ooffset_t writemappings; 1981 int error, nseals; 1982 1983 error = 0; 1984 shmfd = fp->f_data; 1985 rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX); 1986 1987 /* Even already-set seals should result in EPERM. */ 1988 if ((shmfd->shm_seals & F_SEAL_SEAL) != 0) { 1989 error = EPERM; 1990 goto out; 1991 } 1992 nseals = seals & ~shmfd->shm_seals; 1993 if ((nseals & F_SEAL_WRITE) != 0) { 1994 if (shm_largepage(shmfd)) { 1995 error = ENOTSUP; 1996 goto out; 1997 } 1998 1999 /* 2000 * The rangelock above prevents writable mappings from being 2001 * added after we've started applying seals. The RLOCK here 2002 * is to avoid torn reads on ILP32 arches as unmapping/reducing 2003 * writemappings will be done without a rangelock. 2004 */ 2005 VM_OBJECT_RLOCK(shmfd->shm_object); 2006 writemappings = shmfd->shm_object->un_pager.swp.writemappings; 2007 VM_OBJECT_RUNLOCK(shmfd->shm_object); 2008 /* kmappings are also writable */ 2009 if (writemappings > 0) { 2010 error = EBUSY; 2011 goto out; 2012 } 2013 } 2014 shmfd->shm_seals |= nseals; 2015 out: 2016 shm_rangelock_unlock(shmfd, rl_cookie); 2017 return (error); 2018 } 2019 2020 static int 2021 shm_get_seals(struct file *fp, int *seals) 2022 { 2023 struct shmfd *shmfd; 2024 2025 shmfd = fp->f_data; 2026 *seals = shmfd->shm_seals; 2027 return (0); 2028 } 2029 2030 static int 2031 shm_deallocate(struct shmfd *shmfd, off_t *offset, off_t *length, int flags) 2032 { 2033 vm_object_t object; 2034 vm_pindex_t pistart, pi, piend; 2035 vm_ooffset_t off, len; 2036 int startofs, endofs, end; 2037 int error; 2038 2039 off = *offset; 2040 len = *length; 2041 KASSERT(off + len <= (vm_ooffset_t)OFF_MAX, ("off + len overflows")); 2042 if (off + len > shmfd->shm_size) 2043 len = shmfd->shm_size - off; 2044 object = shmfd->shm_object; 2045 startofs = off & PAGE_MASK; 2046 endofs = (off + len) & PAGE_MASK; 2047 pistart = OFF_TO_IDX(off); 2048 piend = OFF_TO_IDX(off + len); 2049 pi = OFF_TO_IDX(off + PAGE_MASK); 2050 error = 0; 2051 2052 /* Handle the case when offset is on or beyond shm size. */ 2053 if ((off_t)len <= 0) { 2054 *length = 0; 2055 return (0); 2056 } 2057 2058 VM_OBJECT_WLOCK(object); 2059 2060 if (startofs != 0) { 2061 end = pistart != piend ? PAGE_SIZE : endofs; 2062 error = shm_partial_page_invalidate(object, pistart, startofs, 2063 end); 2064 if (error) 2065 goto out; 2066 off += end - startofs; 2067 len -= end - startofs; 2068 } 2069 2070 if (pi < piend) { 2071 vm_object_page_remove(object, pi, piend, 0); 2072 off += IDX_TO_OFF(piend - pi); 2073 len -= IDX_TO_OFF(piend - pi); 2074 } 2075 2076 if (endofs != 0 && pistart != piend) { 2077 error = shm_partial_page_invalidate(object, piend, 0, endofs); 2078 if (error) 2079 goto out; 2080 off += endofs; 2081 len -= endofs; 2082 } 2083 2084 out: 2085 VM_OBJECT_WUNLOCK(shmfd->shm_object); 2086 *offset = off; 2087 *length = len; 2088 return (error); 2089 } 2090 2091 static int 2092 shm_fspacectl(struct file *fp, int cmd, off_t *offset, off_t *length, int flags, 2093 struct ucred *active_cred, struct thread *td) 2094 { 2095 void *rl_cookie; 2096 struct shmfd *shmfd; 2097 off_t off, len; 2098 int error; 2099 2100 KASSERT(cmd == SPACECTL_DEALLOC, ("shm_fspacectl: Invalid cmd")); 2101 KASSERT((flags & ~SPACECTL_F_SUPPORTED) == 0, 2102 ("shm_fspacectl: non-zero flags")); 2103 KASSERT(*offset >= 0 && *length > 0 && *length <= OFF_MAX - *offset, 2104 ("shm_fspacectl: offset/length overflow or underflow")); 2105 error = EINVAL; 2106 shmfd = fp->f_data; 2107 off = *offset; 2108 len = *length; 2109 2110 rl_cookie = shm_rangelock_wlock(shmfd, off, off + len); 2111 switch (cmd) { 2112 case SPACECTL_DEALLOC: 2113 if ((shmfd->shm_seals & F_SEAL_WRITE) != 0) { 2114 error = EPERM; 2115 break; 2116 } 2117 error = shm_deallocate(shmfd, &off, &len, flags); 2118 *offset = off; 2119 *length = len; 2120 break; 2121 default: 2122 __assert_unreachable(); 2123 } 2124 shm_rangelock_unlock(shmfd, rl_cookie); 2125 return (error); 2126 } 2127 2128 2129 static int 2130 shm_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td) 2131 { 2132 void *rl_cookie; 2133 struct shmfd *shmfd; 2134 size_t size; 2135 int error; 2136 2137 /* This assumes that the caller already checked for overflow. */ 2138 error = 0; 2139 shmfd = fp->f_data; 2140 size = offset + len; 2141 2142 /* 2143 * Just grab the rangelock for the range that we may be attempting to 2144 * grow, rather than blocking read/write for regions we won't be 2145 * touching while this (potential) resize is in progress. Other 2146 * attempts to resize the shmfd will have to take a write lock from 0 to 2147 * OFF_MAX, so this being potentially beyond the current usable range of 2148 * the shmfd is not necessarily a concern. If other mechanisms are 2149 * added to grow a shmfd, this may need to be re-evaluated. 2150 */ 2151 rl_cookie = shm_rangelock_wlock(shmfd, offset, size); 2152 if (size > shmfd->shm_size) 2153 error = shm_dotruncate_cookie(shmfd, size, rl_cookie); 2154 shm_rangelock_unlock(shmfd, rl_cookie); 2155 /* Translate to posix_fallocate(2) return value as needed. */ 2156 if (error == ENOMEM) 2157 error = ENOSPC; 2158 return (error); 2159 } 2160 2161 static int 2162 sysctl_posix_shm_list(SYSCTL_HANDLER_ARGS) 2163 { 2164 struct shm_mapping *shmm; 2165 struct sbuf sb; 2166 struct kinfo_file kif; 2167 u_long i; 2168 int error, error2; 2169 2170 sbuf_new_for_sysctl(&sb, NULL, sizeof(struct kinfo_file) * 5, req); 2171 sbuf_clear_flags(&sb, SBUF_INCLUDENUL); 2172 error = 0; 2173 sx_slock(&shm_dict_lock); 2174 for (i = 0; i < shm_hash + 1; i++) { 2175 LIST_FOREACH(shmm, &shm_dictionary[i], sm_link) { 2176 error = shm_fill_kinfo_locked(shmm->sm_shmfd, 2177 &kif, true); 2178 if (error == EPERM) { 2179 error = 0; 2180 continue; 2181 } 2182 if (error != 0) 2183 break; 2184 pack_kinfo(&kif); 2185 error = sbuf_bcat(&sb, &kif, kif.kf_structsize) == 0 ? 2186 0 : ENOMEM; 2187 if (error != 0) 2188 break; 2189 } 2190 } 2191 sx_sunlock(&shm_dict_lock); 2192 error2 = sbuf_finish(&sb); 2193 sbuf_delete(&sb); 2194 return (error != 0 ? error : error2); 2195 } 2196 2197 SYSCTL_PROC(_kern_ipc, OID_AUTO, posix_shm_list, 2198 CTLFLAG_RD | CTLFLAG_PRISON | CTLFLAG_MPSAFE | CTLTYPE_OPAQUE, 2199 NULL, 0, sysctl_posix_shm_list, "", 2200 "POSIX SHM list"); 2201 2202 int 2203 kern_shm_open(struct thread *td, const char *path, int flags, mode_t mode, 2204 struct filecaps *caps) 2205 { 2206 2207 return (kern_shm_open2(td, path, flags, mode, 0, caps, NULL)); 2208 } 2209 2210 /* 2211 * This version of the shm_open() interface leaves CLOEXEC behavior up to the 2212 * caller, and libc will enforce it for the traditional shm_open() call. This 2213 * allows other consumers, like memfd_create(), to opt-in for CLOEXEC. This 2214 * interface also includes a 'name' argument that is currently unused, but could 2215 * potentially be exported later via some interface for debugging purposes. 2216 * From the kernel's perspective, it is optional. Individual consumers like 2217 * memfd_create() may require it in order to be compatible with other systems 2218 * implementing the same function. 2219 */ 2220 int 2221 sys_shm_open2(struct thread *td, struct shm_open2_args *uap) 2222 { 2223 2224 return (kern_shm_open2(td, uap->path, uap->flags, uap->mode, 2225 uap->shmflags, NULL, uap->name)); 2226 } 2227 2228 int 2229 shm_get_path(struct vm_object *obj, char *path, size_t sz) 2230 { 2231 struct shmfd *shmfd; 2232 int error; 2233 2234 error = 0; 2235 shmfd = NULL; 2236 sx_slock(&shm_dict_lock); 2237 VM_OBJECT_RLOCK(obj); 2238 if ((obj->flags & OBJ_POSIXSHM) == 0) { 2239 error = EINVAL; 2240 } else { 2241 if (obj->type == shmfd_pager_type) 2242 shmfd = obj->un_pager.swp.swp_priv; 2243 else if (obj->type == OBJT_PHYS) 2244 shmfd = obj->un_pager.phys.phys_priv; 2245 if (shmfd == NULL) { 2246 error = ENXIO; 2247 } else { 2248 strlcpy(path, shmfd->shm_path == NULL ? "anon" : 2249 shmfd->shm_path, sz); 2250 } 2251 } 2252 if (error != 0) 2253 path[0] = '\0'; 2254 VM_OBJECT_RUNLOCK(obj); 2255 sx_sunlock(&shm_dict_lock); 2256 return (error); 2257 } 2258