1 /*- 2 * SPDX-License-Identifier: BSD-3-Clause 3 * 4 * Copyright (c) 1999-2008 Apple Inc. 5 * Copyright (c) 2006-2008, 2016, 2018 Robert N. M. Watson 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 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 1. Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 2. Redistributions in binary form must reproduce the above copyright 19 * notice, this list of conditions and the following disclaimer in the 20 * documentation and/or other materials provided with the distribution. 21 * 3. Neither the name of Apple Inc. ("Apple") nor the names of 22 * its contributors may be used to endorse or promote products derived 23 * from this software without specific prior written permission. 24 * 25 * THIS SOFTWARE IS PROVIDED BY APPLE AND ITS 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 APPLE OR ITS CONTRIBUTORS BE LIABLE FOR 29 * 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, 33 * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING 34 * IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 35 * POSSIBILITY OF SUCH DAMAGE. 36 */ 37 38 #include <sys/cdefs.h> 39 __FBSDID("$FreeBSD$"); 40 41 #include <sys/param.h> 42 #include <sys/condvar.h> 43 #include <sys/conf.h> 44 #include <sys/file.h> 45 #include <sys/filedesc.h> 46 #include <sys/fcntl.h> 47 #include <sys/ipc.h> 48 #include <sys/kernel.h> 49 #include <sys/kthread.h> 50 #include <sys/malloc.h> 51 #include <sys/mount.h> 52 #include <sys/namei.h> 53 #include <sys/proc.h> 54 #include <sys/queue.h> 55 #include <sys/socket.h> 56 #include <sys/socketvar.h> 57 #include <sys/protosw.h> 58 #include <sys/domain.h> 59 #include <sys/sx.h> 60 #include <sys/sysproto.h> 61 #include <sys/sysent.h> 62 #include <sys/systm.h> 63 #include <sys/ucred.h> 64 #include <sys/uio.h> 65 #include <sys/un.h> 66 #include <sys/unistd.h> 67 #include <sys/vnode.h> 68 69 #include <bsm/audit.h> 70 #include <bsm/audit_internal.h> 71 #include <bsm/audit_kevents.h> 72 73 #include <netinet/in.h> 74 #include <netinet/in_pcb.h> 75 76 #include <security/audit/audit.h> 77 #include <security/audit/audit_private.h> 78 79 #include <vm/uma.h> 80 81 #include <machine/stdarg.h> 82 83 /* 84 * Worker thread that will schedule disk I/O, etc. 85 */ 86 static struct proc *audit_thread; 87 88 /* 89 * audit_cred and audit_vp are the stored credential and vnode to use for 90 * active audit trail. They are protected by the audit worker lock, which 91 * will be held across all I/O and all rotation to prevent them from being 92 * replaced (rotated) while in use. The audit_file_rotate_wait flag is set 93 * when the kernel has delivered a trigger to auditd to rotate the trail, and 94 * is cleared when the next rotation takes place. It is also protected by 95 * the audit worker lock. 96 */ 97 static int audit_file_rotate_wait; 98 static struct ucred *audit_cred; 99 static struct vnode *audit_vp; 100 static off_t audit_size; 101 static struct sx audit_worker_lock; 102 103 #define AUDIT_WORKER_LOCK_INIT() sx_init(&audit_worker_lock, \ 104 "audit_worker_lock"); 105 #define AUDIT_WORKER_LOCK_ASSERT() sx_assert(&audit_worker_lock, \ 106 SA_XLOCKED) 107 #define AUDIT_WORKER_LOCK() sx_xlock(&audit_worker_lock) 108 #define AUDIT_WORKER_UNLOCK() sx_xunlock(&audit_worker_lock) 109 110 static void 111 audit_worker_sync_vp(struct vnode *vp, struct mount *mp, const char *fmt, ...) 112 { 113 struct mount *mp1; 114 int error; 115 va_list va; 116 117 va_start(va, fmt); 118 error = vn_start_write(vp, &mp1, 0); 119 if (error == 0) { 120 VOP_LOCK(vp, LK_EXCLUSIVE | LK_RETRY); 121 (void)VOP_FSYNC(vp, MNT_WAIT, curthread); 122 VOP_UNLOCK(vp); 123 vn_finished_write(mp1); 124 } 125 vfs_unbusy(mp); 126 vpanic(fmt, va); 127 va_end(va); 128 } 129 130 /* 131 * Write an audit record to a file, performed as the last stage after both 132 * preselection and BSM conversion. Both space management and write failures 133 * are handled in this function. 134 * 135 * No attempt is made to deal with possible failure to deliver a trigger to 136 * the audit daemon, since the message is asynchronous anyway. 137 */ 138 static void 139 audit_record_write(struct vnode *vp, struct ucred *cred, void *data, 140 size_t len) 141 { 142 static struct timeval last_lowspace_trigger; 143 static struct timeval last_fail; 144 static int cur_lowspace_trigger; 145 struct statfs *mnt_stat; 146 struct mount *mp; 147 int error; 148 static int cur_fail; 149 long temp; 150 151 AUDIT_WORKER_LOCK_ASSERT(); 152 153 if (vp == NULL) 154 return; 155 156 mp = vp->v_mount; 157 if (mp == NULL) { 158 error = EINVAL; 159 goto fail; 160 } 161 error = vfs_busy(mp, 0); 162 if (error != 0) { 163 mp = NULL; 164 goto fail; 165 } 166 mnt_stat = &mp->mnt_stat; 167 168 /* 169 * First, gather statistics on the audit log file and file system so 170 * that we know how we're doing on space. Consider failure of these 171 * operations to indicate a future inability to write to the file. 172 */ 173 error = VFS_STATFS(mp, mnt_stat); 174 if (error != 0) 175 goto fail; 176 177 /* 178 * We handle four different space-related limits: 179 * 180 * - A fixed (hard) limit on the minimum free blocks we require on 181 * the file system, and results in record loss, a trigger, and 182 * possible fail stop due to violating invariants. 183 * 184 * - An administrative (soft) limit, which when fallen below, results 185 * in the kernel notifying the audit daemon of low space. 186 * 187 * - An audit trail size limit, which when gone above, results in the 188 * kernel notifying the audit daemon that rotation is desired. 189 * 190 * - The total depth of the kernel audit record exceeding free space, 191 * which can lead to possible fail stop (with drain), in order to 192 * prevent violating invariants. Failure here doesn't halt 193 * immediately, but prevents new records from being generated. 194 * 195 * Possibly, the last of these should be handled differently, always 196 * allowing a full queue to be lost, rather than trying to prevent 197 * loss. 198 * 199 * First, handle the hard limit, which generates a trigger and may 200 * fail stop. This is handled in the same manner as ENOSPC from 201 * VOP_WRITE, and results in record loss. 202 */ 203 if (mnt_stat->f_bfree < AUDIT_HARD_LIMIT_FREE_BLOCKS) { 204 error = ENOSPC; 205 goto fail_enospc; 206 } 207 208 /* 209 * Second, handle falling below the soft limit, if defined; we send 210 * the daemon a trigger and continue processing the record. Triggers 211 * are limited to 1/sec. 212 */ 213 if (audit_qctrl.aq_minfree != 0) { 214 temp = mnt_stat->f_blocks / (100 / audit_qctrl.aq_minfree); 215 if (mnt_stat->f_bfree < temp) { 216 if (ppsratecheck(&last_lowspace_trigger, 217 &cur_lowspace_trigger, 1)) { 218 (void)audit_send_trigger( 219 AUDIT_TRIGGER_LOW_SPACE); 220 printf("Warning: disk space low (< %d%% free) " 221 "on audit log file-system\n", 222 audit_qctrl.aq_minfree); 223 } 224 } 225 } 226 227 /* 228 * If the current file is getting full, generate a rotation trigger 229 * to the daemon. This is only approximate, which is fine as more 230 * records may be generated before the daemon rotates the file. 231 */ 232 if (audit_fstat.af_filesz != 0 && 233 audit_size >= audit_fstat.af_filesz * (audit_file_rotate_wait + 1)) { 234 AUDIT_WORKER_LOCK_ASSERT(); 235 236 audit_file_rotate_wait++; 237 (void)audit_send_trigger(AUDIT_TRIGGER_ROTATE_KERNEL); 238 } 239 240 /* 241 * If the estimated amount of audit data in the audit event queue 242 * (plus records allocated but not yet queued) has reached the amount 243 * of free space on the disk, then we need to go into an audit fail 244 * stop state, in which we do not permit the allocation/committing of 245 * any new audit records. We continue to process records but don't 246 * allow any activities that might generate new records. In the 247 * future, we might want to detect when space is available again and 248 * allow operation to continue, but this behavior is sufficient to 249 * meet fail stop requirements in CAPP. 250 */ 251 if (audit_fail_stop) { 252 if ((unsigned long)((audit_q_len + audit_pre_q_len + 1) * 253 MAX_AUDIT_RECORD_SIZE) / mnt_stat->f_bsize >= 254 (unsigned long)(mnt_stat->f_bfree)) { 255 if (ppsratecheck(&last_fail, &cur_fail, 1)) 256 printf("audit_record_write: free space " 257 "below size of audit queue, failing " 258 "stop\n"); 259 audit_in_failure = 1; 260 } else if (audit_in_failure) { 261 /* 262 * Note: if we want to handle recovery, this is the 263 * spot to do it: unset audit_in_failure, and issue a 264 * wakeup on the cv. 265 */ 266 } 267 } 268 269 error = vn_rdwr(UIO_WRITE, vp, data, len, (off_t)0, UIO_SYSSPACE, 270 IO_APPEND|IO_UNIT, cred, NULL, NULL, curthread); 271 if (error == ENOSPC) 272 goto fail_enospc; 273 else if (error) 274 goto fail; 275 AUDIT_WORKER_LOCK_ASSERT(); 276 audit_size += len; 277 278 /* 279 * Catch completion of a queue drain here; if we're draining and the 280 * queue is now empty, fail stop. That audit_fail_stop is implicitly 281 * true, since audit_in_failure can only be set of audit_fail_stop is 282 * set. 283 * 284 * Note: if we handle recovery from audit_in_failure, then we need to 285 * make panic here conditional. 286 */ 287 if (audit_in_failure) { 288 if (audit_q_len == 0 && audit_pre_q_len == 0) { 289 audit_worker_sync_vp(vp, mp, 290 "Audit store overflow; record queue drained."); 291 } 292 } 293 294 vfs_unbusy(mp); 295 return; 296 297 fail_enospc: 298 /* 299 * ENOSPC is considered a special case with respect to failures, as 300 * this can reflect either our preemptive detection of insufficient 301 * space, or ENOSPC returned by the vnode write call. 302 */ 303 if (audit_fail_stop) { 304 audit_worker_sync_vp(vp, mp, 305 "Audit log space exhausted and fail-stop set."); 306 } 307 (void)audit_send_trigger(AUDIT_TRIGGER_NO_SPACE); 308 audit_trail_suspended = 1; 309 audit_syscalls_enabled_update(); 310 311 /* FALLTHROUGH */ 312 fail: 313 /* 314 * We have failed to write to the file, so the current record is 315 * lost, which may require an immediate system halt. 316 */ 317 if (audit_panic_on_write_fail) { 318 audit_worker_sync_vp(vp, mp, 319 "audit_worker: write error %d\n", error); 320 } else if (ppsratecheck(&last_fail, &cur_fail, 1)) 321 printf("audit_worker: write error %d\n", error); 322 if (mp != NULL) 323 vfs_unbusy(mp); 324 } 325 326 /* 327 * Given a kernel audit record, process as required. Kernel audit records 328 * are converted to one, or possibly two, BSM records, depending on whether 329 * there is a user audit record present also. Kernel records need be 330 * converted to BSM before they can be written out. Both types will be 331 * written to disk, and audit pipes. 332 */ 333 static void 334 audit_worker_process_record(struct kaudit_record *ar) 335 { 336 struct au_record *bsm; 337 au_class_t class; 338 au_event_t event; 339 au_id_t auid; 340 int error, sorf; 341 int locked; 342 343 /* 344 * We hold the audit worker lock over both writes, if there are two, 345 * so that the two records won't be split across a rotation and end 346 * up in two different trail files. 347 */ 348 if (((ar->k_ar_commit & AR_COMMIT_USER) && 349 (ar->k_ar_commit & AR_PRESELECT_USER_TRAIL)) || 350 (ar->k_ar_commit & AR_PRESELECT_TRAIL)) { 351 AUDIT_WORKER_LOCK(); 352 locked = 1; 353 } else 354 locked = 0; 355 356 /* 357 * First, handle the user record, if any: commit to the system trail 358 * and audit pipes as selected. 359 */ 360 if ((ar->k_ar_commit & AR_COMMIT_USER) && 361 (ar->k_ar_commit & AR_PRESELECT_USER_TRAIL)) { 362 AUDIT_WORKER_LOCK_ASSERT(); 363 audit_record_write(audit_vp, audit_cred, ar->k_udata, 364 ar->k_ulen); 365 } 366 367 if ((ar->k_ar_commit & AR_COMMIT_USER) && 368 (ar->k_ar_commit & AR_PRESELECT_USER_PIPE)) 369 audit_pipe_submit_user(ar->k_udata, ar->k_ulen); 370 371 if (!(ar->k_ar_commit & AR_COMMIT_KERNEL) || 372 ((ar->k_ar_commit & AR_PRESELECT_PIPE) == 0 && 373 (ar->k_ar_commit & AR_PRESELECT_TRAIL) == 0 && 374 (ar->k_ar_commit & AR_PRESELECT_DTRACE) == 0)) 375 goto out; 376 377 auid = ar->k_ar.ar_subj_auid; 378 event = ar->k_ar.ar_event; 379 class = au_event_class(event); 380 if (ar->k_ar.ar_errno == 0) 381 sorf = AU_PRS_SUCCESS; 382 else 383 sorf = AU_PRS_FAILURE; 384 385 error = kaudit_to_bsm(ar, &bsm); 386 switch (error) { 387 case BSM_NOAUDIT: 388 goto out; 389 390 case BSM_FAILURE: 391 printf("audit_worker_process_record: BSM_FAILURE\n"); 392 goto out; 393 394 case BSM_SUCCESS: 395 break; 396 397 default: 398 panic("kaudit_to_bsm returned %d", error); 399 } 400 401 if (ar->k_ar_commit & AR_PRESELECT_TRAIL) { 402 AUDIT_WORKER_LOCK_ASSERT(); 403 audit_record_write(audit_vp, audit_cred, bsm->data, bsm->len); 404 } 405 406 if (ar->k_ar_commit & AR_PRESELECT_PIPE) 407 audit_pipe_submit(auid, event, class, sorf, 408 ar->k_ar_commit & AR_PRESELECT_TRAIL, bsm->data, 409 bsm->len); 410 411 #ifdef KDTRACE_HOOKS 412 /* 413 * Version of the dtaudit commit hook that accepts BSM. 414 */ 415 if (ar->k_ar_commit & AR_PRESELECT_DTRACE) { 416 if (dtaudit_hook_bsm != NULL) 417 dtaudit_hook_bsm(ar, auid, event, class, sorf, 418 bsm->data, bsm->len); 419 } 420 #endif 421 422 kau_free(bsm); 423 out: 424 if (locked) 425 AUDIT_WORKER_UNLOCK(); 426 } 427 428 /* 429 * The audit_worker thread is responsible for watching the event queue, 430 * dequeueing records, converting them to BSM format, and committing them to 431 * disk. In order to minimize lock thrashing, records are dequeued in sets 432 * to a thread-local work queue. 433 * 434 * Note: this means that the effect bound on the size of the pending record 435 * queue is 2x the length of the global queue. 436 */ 437 static void 438 audit_worker(void *arg) 439 { 440 struct kaudit_queue ar_worklist; 441 struct kaudit_record *ar; 442 int lowater_signal; 443 444 TAILQ_INIT(&ar_worklist); 445 mtx_lock(&audit_mtx); 446 while (1) { 447 mtx_assert(&audit_mtx, MA_OWNED); 448 449 /* 450 * Wait for a record. 451 */ 452 while (TAILQ_EMPTY(&audit_q)) 453 cv_wait(&audit_worker_cv, &audit_mtx); 454 455 /* 456 * If there are records in the global audit record queue, 457 * transfer them to a thread-local queue and process them 458 * one by one. If we cross the low watermark threshold, 459 * signal any waiting processes that they may wake up and 460 * continue generating records. 461 */ 462 lowater_signal = 0; 463 while ((ar = TAILQ_FIRST(&audit_q))) { 464 TAILQ_REMOVE(&audit_q, ar, k_q); 465 audit_q_len--; 466 if (audit_q_len == audit_qctrl.aq_lowater) 467 lowater_signal++; 468 TAILQ_INSERT_TAIL(&ar_worklist, ar, k_q); 469 } 470 if (lowater_signal) 471 cv_broadcast(&audit_watermark_cv); 472 473 mtx_unlock(&audit_mtx); 474 while ((ar = TAILQ_FIRST(&ar_worklist))) { 475 TAILQ_REMOVE(&ar_worklist, ar, k_q); 476 audit_worker_process_record(ar); 477 audit_free(ar); 478 } 479 mtx_lock(&audit_mtx); 480 } 481 } 482 483 /* 484 * audit_rotate_vnode() is called by a user or kernel thread to configure or 485 * de-configure auditing on a vnode. The arguments are the replacement 486 * credential (referenced) and vnode (referenced and opened) to substitute 487 * for the current credential and vnode, if any. If either is set to NULL, 488 * both should be NULL, and this is used to indicate that audit is being 489 * disabled. Any previous cred/vnode will be closed and freed. We re-enable 490 * generating rotation requests to auditd. 491 */ 492 void 493 audit_rotate_vnode(struct ucred *cred, struct vnode *vp) 494 { 495 struct ucred *old_audit_cred; 496 struct vnode *old_audit_vp; 497 struct vattr vattr; 498 499 KASSERT((cred != NULL && vp != NULL) || (cred == NULL && vp == NULL), 500 ("audit_rotate_vnode: cred %p vp %p", cred, vp)); 501 502 if (vp != NULL) { 503 vn_lock(vp, LK_SHARED | LK_RETRY); 504 if (VOP_GETATTR(vp, &vattr, cred) != 0) 505 vattr.va_size = 0; 506 VOP_UNLOCK(vp); 507 } else { 508 vattr.va_size = 0; 509 } 510 511 /* 512 * Rotate the vnode/cred, and clear the rotate flag so that we will 513 * send a rotate trigger if the new file fills. 514 */ 515 AUDIT_WORKER_LOCK(); 516 old_audit_cred = audit_cred; 517 old_audit_vp = audit_vp; 518 audit_cred = cred; 519 audit_vp = vp; 520 audit_size = vattr.va_size; 521 audit_file_rotate_wait = 0; 522 audit_trail_enabled = (audit_vp != NULL); 523 audit_syscalls_enabled_update(); 524 AUDIT_WORKER_UNLOCK(); 525 526 /* 527 * If there was an old vnode/credential, close and free. 528 */ 529 if (old_audit_vp != NULL) { 530 vn_close(old_audit_vp, AUDIT_CLOSE_FLAGS, old_audit_cred, 531 curthread); 532 crfree(old_audit_cred); 533 } 534 } 535 536 void 537 audit_worker_init(void) 538 { 539 int error; 540 541 AUDIT_WORKER_LOCK_INIT(); 542 error = kproc_create(audit_worker, NULL, &audit_thread, RFHIGHPID, 543 0, "audit"); 544 if (error) 545 panic("audit_worker_init: kproc_create returned %d", error); 546 } 547