1 /* 2 * Copyright (c) 1999-2005 Apple Computer, Inc. 3 * Copyright (c) 2006-2008 Robert N. M. Watson 4 * All rights reserved. 5 * 6 * Redistribution and use in source and binary forms, with or without 7 * modification, are permitted provided that the following conditions 8 * are met: 9 * 1. Redistributions of source code must retain the above copyright 10 * notice, this list of conditions and the following disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 3. Neither the name of Apple Computer, Inc. ("Apple") nor the names of 15 * its contributors may be used to endorse or promote products derived 16 * from this software without specific prior written permission. 17 * 18 * THIS SOFTWARE IS PROVIDED BY APPLE AND ITS CONTRIBUTORS "AS IS" AND 19 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 20 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 21 * ARE DISCLAIMED. IN NO EVENT SHALL APPLE OR ITS CONTRIBUTORS BE LIABLE FOR 22 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 23 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 24 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 25 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, 26 * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING 27 * IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 28 * POSSIBILITY OF SUCH DAMAGE. 29 */ 30 31 #include <sys/cdefs.h> 32 __FBSDID("$FreeBSD$"); 33 34 #include <sys/param.h> 35 #include <sys/condvar.h> 36 #include <sys/conf.h> 37 #include <sys/file.h> 38 #include <sys/filedesc.h> 39 #include <sys/fcntl.h> 40 #include <sys/ipc.h> 41 #include <sys/kernel.h> 42 #include <sys/kthread.h> 43 #include <sys/malloc.h> 44 #include <sys/mount.h> 45 #include <sys/namei.h> 46 #include <sys/proc.h> 47 #include <sys/queue.h> 48 #include <sys/socket.h> 49 #include <sys/socketvar.h> 50 #include <sys/protosw.h> 51 #include <sys/domain.h> 52 #include <sys/sx.h> 53 #include <sys/sysproto.h> 54 #include <sys/sysent.h> 55 #include <sys/systm.h> 56 #include <sys/ucred.h> 57 #include <sys/uio.h> 58 #include <sys/un.h> 59 #include <sys/unistd.h> 60 #include <sys/vnode.h> 61 62 #include <bsm/audit.h> 63 #include <bsm/audit_internal.h> 64 #include <bsm/audit_kevents.h> 65 66 #include <netinet/in.h> 67 #include <netinet/in_pcb.h> 68 69 #include <security/audit/audit.h> 70 #include <security/audit/audit_private.h> 71 72 #include <vm/uma.h> 73 74 /* 75 * Worker thread that will schedule disk I/O, etc. 76 */ 77 static struct proc *audit_thread; 78 79 /* 80 * audit_cred and audit_vp are the stored credential and vnode to use for 81 * active audit trail. They are protected by audit_worker_sx, which will be 82 * held across all I/O and all rotation to prevent them from being replaced 83 * (rotated) while in use. The audit_file_rotate_wait flag is set when the 84 * kernel has delivered a trigger to auditd to rotate the trail, and is 85 * cleared when the next rotation takes place. It is also protected by 86 * audit_worker_sx. 87 */ 88 static int audit_file_rotate_wait; 89 static struct sx audit_worker_sx; 90 static struct ucred *audit_cred; 91 static struct vnode *audit_vp; 92 93 /* 94 * Write an audit record to a file, performed as the last stage after both 95 * preselection and BSM conversion. Both space management and write failures 96 * are handled in this function. 97 * 98 * No attempt is made to deal with possible failure to deliver a trigger to 99 * the audit daemon, since the message is asynchronous anyway. 100 */ 101 static void 102 audit_record_write(struct vnode *vp, struct ucred *cred, void *data, 103 size_t len) 104 { 105 static struct timeval last_lowspace_trigger; 106 static struct timeval last_fail; 107 static int cur_lowspace_trigger; 108 struct statfs *mnt_stat; 109 int error, vfslocked; 110 static int cur_fail; 111 struct vattr vattr; 112 long temp; 113 114 sx_assert(&audit_worker_sx, SA_LOCKED); /* audit_file_rotate_wait. */ 115 116 if (vp == NULL) 117 return; 118 119 mnt_stat = &vp->v_mount->mnt_stat; 120 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 121 122 /* 123 * First, gather statistics on the audit log file and file system so 124 * that we know how we're doing on space. Consider failure of these 125 * operations to indicate a future inability to write to the file. 126 */ 127 error = VFS_STATFS(vp->v_mount, mnt_stat, curthread); 128 if (error) 129 goto fail; 130 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 131 error = VOP_GETATTR(vp, &vattr, cred, curthread); 132 VOP_UNLOCK(vp, 0); 133 if (error) 134 goto fail; 135 audit_fstat.af_currsz = vattr.va_size; 136 137 /* 138 * We handle four different space-related limits: 139 * 140 * - A fixed (hard) limit on the minimum free blocks we require on 141 * the file system, and results in record loss, a trigger, and 142 * possible fail stop due to violating invariants. 143 * 144 * - An administrative (soft) limit, which when fallen below, results 145 * in the kernel notifying the audit daemon of low space. 146 * 147 * - An audit trail size limit, which when gone above, results in the 148 * kernel notifying the audit daemon that rotation is desired. 149 * 150 * - The total depth of the kernel audit record exceeding free space, 151 * which can lead to possible fail stop (with drain), in order to 152 * prevent violating invariants. Failure here doesn't halt 153 * immediately, but prevents new records from being generated. 154 * 155 * Possibly, the last of these should be handled differently, always 156 * allowing a full queue to be lost, rather than trying to prevent 157 * loss. 158 * 159 * First, handle the hard limit, which generates a trigger and may 160 * fail stop. This is handled in the same manner as ENOSPC from 161 * VOP_WRITE, and results in record loss. 162 */ 163 if (mnt_stat->f_bfree < AUDIT_HARD_LIMIT_FREE_BLOCKS) { 164 error = ENOSPC; 165 goto fail_enospc; 166 } 167 168 /* 169 * Second, handle falling below the soft limit, if defined; we send 170 * the daemon a trigger and continue processing the record. Triggers 171 * are limited to 1/sec. 172 */ 173 if (audit_qctrl.aq_minfree != 0) { 174 temp = mnt_stat->f_blocks / (100 / audit_qctrl.aq_minfree); 175 if (mnt_stat->f_bfree < temp) { 176 if (ppsratecheck(&last_lowspace_trigger, 177 &cur_lowspace_trigger, 1)) { 178 (void)audit_send_trigger( 179 AUDIT_TRIGGER_LOW_SPACE); 180 printf("Warning: audit space low\n"); 181 } 182 } 183 } 184 185 /* 186 * If the current file is getting full, generate a rotation trigger 187 * to the daemon. This is only approximate, which is fine as more 188 * records may be generated before the daemon rotates the file. 189 */ 190 if ((audit_fstat.af_filesz != 0) && (audit_file_rotate_wait == 0) && 191 (vattr.va_size >= audit_fstat.af_filesz)) { 192 sx_assert(&audit_worker_sx, SA_XLOCKED); 193 194 audit_file_rotate_wait = 1; 195 (void)audit_send_trigger(AUDIT_TRIGGER_ROTATE_KERNEL); 196 } 197 198 /* 199 * If the estimated amount of audit data in the audit event queue 200 * (plus records allocated but not yet queued) has reached the amount 201 * of free space on the disk, then we need to go into an audit fail 202 * stop state, in which we do not permit the allocation/committing of 203 * any new audit records. We continue to process records but don't 204 * allow any activities that might generate new records. In the 205 * future, we might want to detect when space is available again and 206 * allow operation to continue, but this behavior is sufficient to 207 * meet fail stop requirements in CAPP. 208 */ 209 if (audit_fail_stop) { 210 if ((unsigned long)((audit_q_len + audit_pre_q_len + 1) * 211 MAX_AUDIT_RECORD_SIZE) / mnt_stat->f_bsize >= 212 (unsigned long)(mnt_stat->f_bfree)) { 213 if (ppsratecheck(&last_fail, &cur_fail, 1)) 214 printf("audit_record_write: free space " 215 "below size of audit queue, failing " 216 "stop\n"); 217 audit_in_failure = 1; 218 } else if (audit_in_failure) { 219 /* 220 * Note: if we want to handle recovery, this is the 221 * spot to do it: unset audit_in_failure, and issue a 222 * wakeup on the cv. 223 */ 224 } 225 } 226 227 error = vn_rdwr(UIO_WRITE, vp, data, len, (off_t)0, UIO_SYSSPACE, 228 IO_APPEND|IO_UNIT, cred, NULL, NULL, curthread); 229 if (error == ENOSPC) 230 goto fail_enospc; 231 else if (error) 232 goto fail; 233 234 /* 235 * Catch completion of a queue drain here; if we're draining and the 236 * queue is now empty, fail stop. That audit_fail_stop is implicitly 237 * true, since audit_in_failure can only be set of audit_fail_stop is 238 * set. 239 * 240 * Note: if we handle recovery from audit_in_failure, then we need to 241 * make panic here conditional. 242 */ 243 if (audit_in_failure) { 244 if (audit_q_len == 0 && audit_pre_q_len == 0) { 245 VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK); 246 (void)VOP_FSYNC(vp, MNT_WAIT, curthread); 247 VOP_UNLOCK(vp, 0); 248 panic("Audit store overflow; record queue drained."); 249 } 250 } 251 252 VFS_UNLOCK_GIANT(vfslocked); 253 return; 254 255 fail_enospc: 256 /* 257 * ENOSPC is considered a special case with respect to failures, as 258 * this can reflect either our preemptive detection of insufficient 259 * space, or ENOSPC returned by the vnode write call. 260 */ 261 if (audit_fail_stop) { 262 VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK); 263 (void)VOP_FSYNC(vp, MNT_WAIT, curthread); 264 VOP_UNLOCK(vp, 0); 265 panic("Audit log space exhausted and fail-stop set."); 266 } 267 (void)audit_send_trigger(AUDIT_TRIGGER_NO_SPACE); 268 audit_suspended = 1; 269 270 /* FALLTHROUGH */ 271 fail: 272 /* 273 * We have failed to write to the file, so the current record is 274 * lost, which may require an immediate system halt. 275 */ 276 if (audit_panic_on_write_fail) { 277 VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK); 278 (void)VOP_FSYNC(vp, MNT_WAIT, curthread); 279 VOP_UNLOCK(vp, 0); 280 panic("audit_worker: write error %d\n", error); 281 } else if (ppsratecheck(&last_fail, &cur_fail, 1)) 282 printf("audit_worker: write error %d\n", error); 283 VFS_UNLOCK_GIANT(vfslocked); 284 } 285 286 /* 287 * Given a kernel audit record, process as required. Kernel audit records 288 * are converted to one, or possibly two, BSM records, depending on whether 289 * there is a user audit record present also. Kernel records need be 290 * converted to BSM before they can be written out. Both types will be 291 * written to disk, and audit pipes. 292 */ 293 static void 294 audit_worker_process_record(struct kaudit_record *ar) 295 { 296 struct au_record *bsm; 297 au_class_t class; 298 au_event_t event; 299 au_id_t auid; 300 int error, sorf; 301 int trail_locked; 302 303 /* 304 * We hold the audit_worker_sx lock over both writes, if there are 305 * two, so that the two records won't be split across a rotation and 306 * end up in two different trail files. 307 */ 308 if (((ar->k_ar_commit & AR_COMMIT_USER) && 309 (ar->k_ar_commit & AR_PRESELECT_USER_TRAIL)) || 310 (ar->k_ar_commit & AR_PRESELECT_TRAIL)) { 311 sx_xlock(&audit_worker_sx); 312 trail_locked = 1; 313 } else 314 trail_locked = 0; 315 316 /* 317 * First, handle the user record, if any: commit to the system trail 318 * and audit pipes as selected. 319 */ 320 if ((ar->k_ar_commit & AR_COMMIT_USER) && 321 (ar->k_ar_commit & AR_PRESELECT_USER_TRAIL)) { 322 sx_assert(&audit_worker_sx, SA_XLOCKED); 323 audit_record_write(audit_vp, audit_cred, ar->k_udata, 324 ar->k_ulen); 325 } 326 327 if ((ar->k_ar_commit & AR_COMMIT_USER) && 328 (ar->k_ar_commit & AR_PRESELECT_USER_PIPE)) 329 audit_pipe_submit_user(ar->k_udata, ar->k_ulen); 330 331 if (!(ar->k_ar_commit & AR_COMMIT_KERNEL) || 332 ((ar->k_ar_commit & AR_PRESELECT_PIPE) == 0 && 333 (ar->k_ar_commit & AR_PRESELECT_TRAIL) == 0)) 334 goto out; 335 336 auid = ar->k_ar.ar_subj_auid; 337 event = ar->k_ar.ar_event; 338 class = au_event_class(event); 339 if (ar->k_ar.ar_errno == 0) 340 sorf = AU_PRS_SUCCESS; 341 else 342 sorf = AU_PRS_FAILURE; 343 344 error = kaudit_to_bsm(ar, &bsm); 345 switch (error) { 346 case BSM_NOAUDIT: 347 goto out; 348 349 case BSM_FAILURE: 350 printf("audit_worker_process_record: BSM_FAILURE\n"); 351 goto out; 352 353 case BSM_SUCCESS: 354 break; 355 356 default: 357 panic("kaudit_to_bsm returned %d", error); 358 } 359 360 if (ar->k_ar_commit & AR_PRESELECT_TRAIL) { 361 sx_assert(&audit_worker_sx, SA_XLOCKED); 362 audit_record_write(audit_vp, audit_cred, bsm->data, bsm->len); 363 } 364 365 if (ar->k_ar_commit & AR_PRESELECT_PIPE) 366 audit_pipe_submit(auid, event, class, sorf, 367 ar->k_ar_commit & AR_PRESELECT_TRAIL, bsm->data, 368 bsm->len); 369 370 kau_free(bsm); 371 out: 372 if (trail_locked) 373 sx_xunlock(&audit_worker_sx); 374 } 375 376 /* 377 * The audit_worker thread is responsible for watching the event queue, 378 * dequeueing records, converting them to BSM format, and committing them to 379 * disk. In order to minimize lock thrashing, records are dequeued in sets 380 * to a thread-local work queue. 381 * 382 * Note: this means that the effect bound on the size of the pending record 383 * queue is 2x the length of the global queue. 384 */ 385 static void 386 audit_worker(void *arg) 387 { 388 struct kaudit_queue ar_worklist; 389 struct kaudit_record *ar; 390 int lowater_signal; 391 392 TAILQ_INIT(&ar_worklist); 393 mtx_lock(&audit_mtx); 394 while (1) { 395 mtx_assert(&audit_mtx, MA_OWNED); 396 397 /* 398 * Wait for a record. 399 */ 400 while (TAILQ_EMPTY(&audit_q)) 401 cv_wait(&audit_worker_cv, &audit_mtx); 402 403 /* 404 * If there are records in the global audit record queue, 405 * transfer them to a thread-local queue and process them 406 * one by one. If we cross the low watermark threshold, 407 * signal any waiting processes that they may wake up and 408 * continue generating records. 409 */ 410 lowater_signal = 0; 411 while ((ar = TAILQ_FIRST(&audit_q))) { 412 TAILQ_REMOVE(&audit_q, ar, k_q); 413 audit_q_len--; 414 if (audit_q_len == audit_qctrl.aq_lowater) 415 lowater_signal++; 416 TAILQ_INSERT_TAIL(&ar_worklist, ar, k_q); 417 } 418 if (lowater_signal) 419 cv_broadcast(&audit_watermark_cv); 420 421 mtx_unlock(&audit_mtx); 422 while ((ar = TAILQ_FIRST(&ar_worklist))) { 423 TAILQ_REMOVE(&ar_worklist, ar, k_q); 424 audit_worker_process_record(ar); 425 audit_free(ar); 426 } 427 mtx_lock(&audit_mtx); 428 } 429 } 430 431 /* 432 * audit_rotate_vnode() is called by a user or kernel thread to configure or 433 * de-configure auditing on a vnode. The arguments are the replacement 434 * credential (referenced) and vnode (referenced and opened) to substitute 435 * for the current credential and vnode, if any. If either is set to NULL, 436 * both should be NULL, and this is used to indicate that audit is being 437 * disabled. Any previous cred/vnode will be closed and freed. We re-enable 438 * generating rotation requests to auditd. 439 */ 440 void 441 audit_rotate_vnode(struct ucred *cred, struct vnode *vp) 442 { 443 struct ucred *old_audit_cred; 444 struct vnode *old_audit_vp; 445 int vfslocked; 446 447 KASSERT((cred != NULL && vp != NULL) || (cred == NULL && vp == NULL), 448 ("audit_rotate_vnode: cred %p vp %p", cred, vp)); 449 450 /* 451 * Rotate the vnode/cred, and clear the rotate flag so that we will 452 * send a rotate trigger if the new file fills. 453 */ 454 sx_xlock(&audit_worker_sx); 455 old_audit_cred = audit_cred; 456 old_audit_vp = audit_vp; 457 audit_cred = cred; 458 audit_vp = vp; 459 audit_file_rotate_wait = 0; 460 audit_enabled = (audit_vp != NULL); 461 sx_xunlock(&audit_worker_sx); 462 463 /* 464 * If there was an old vnode/credential, close and free. 465 */ 466 if (old_audit_vp != NULL) { 467 vfslocked = VFS_LOCK_GIANT(old_audit_vp->v_mount); 468 vn_close(old_audit_vp, AUDIT_CLOSE_FLAGS, old_audit_cred, 469 curthread); 470 VFS_UNLOCK_GIANT(vfslocked); 471 crfree(old_audit_cred); 472 } 473 } 474 475 void 476 audit_worker_init(void) 477 { 478 int error; 479 480 sx_init(&audit_worker_sx, "audit_worker_sx"); 481 error = kproc_create(audit_worker, NULL, &audit_thread, RFHIGHPID, 482 0, "audit"); 483 if (error) 484 panic("audit_worker_init: kproc_create returned %d", error); 485 } 486