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