xref: /freebsd/sys/security/audit/audit_worker.c (revision 7cd2dcf07629713e5a3d60472cfe4701b705a167)
1 /*-
2  * Copyright (c) 1999-2008 Apple 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 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 the audit worker lock, which
82  * will be held across all I/O and all rotation to prevent them from being
83  * replaced (rotated) while in use.  The audit_file_rotate_wait flag is set
84  * when the kernel has delivered a trigger to auditd to rotate the trail, and
85  * is cleared when the next rotation takes place.  It is also protected by
86  * the audit worker lock.
87  */
88 static int		 audit_file_rotate_wait;
89 static struct ucred	*audit_cred;
90 static struct vnode	*audit_vp;
91 static struct sx	 audit_worker_lock;
92 
93 #define	AUDIT_WORKER_LOCK_INIT()	sx_init(&audit_worker_lock, \
94 					    "audit_worker_lock");
95 #define	AUDIT_WORKER_LOCK_ASSERT()	sx_assert(&audit_worker_lock, \
96 					    SA_XLOCKED)
97 #define	AUDIT_WORKER_LOCK()		sx_xlock(&audit_worker_lock)
98 #define	AUDIT_WORKER_UNLOCK()		sx_xunlock(&audit_worker_lock)
99 
100 /*
101  * Write an audit record to a file, performed as the last stage after both
102  * preselection and BSM conversion.  Both space management and write failures
103  * are handled in this function.
104  *
105  * No attempt is made to deal with possible failure to deliver a trigger to
106  * the audit daemon, since the message is asynchronous anyway.
107  */
108 static void
109 audit_record_write(struct vnode *vp, struct ucred *cred, void *data,
110     size_t len)
111 {
112 	static struct timeval last_lowspace_trigger;
113 	static struct timeval last_fail;
114 	static int cur_lowspace_trigger;
115 	struct statfs *mnt_stat;
116 	int error;
117 	static int cur_fail;
118 	struct vattr vattr;
119 	long temp;
120 
121 	AUDIT_WORKER_LOCK_ASSERT();
122 
123 	if (vp == NULL)
124 		return;
125 
126 	mnt_stat = &vp->v_mount->mnt_stat;
127 
128 	/*
129 	 * First, gather statistics on the audit log file and file system so
130 	 * that we know how we're doing on space.  Consider failure of these
131 	 * operations to indicate a future inability to write to the file.
132 	 */
133 	error = VFS_STATFS(vp->v_mount, mnt_stat);
134 	if (error)
135 		goto fail;
136 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
137 	error = VOP_GETATTR(vp, &vattr, cred);
138 	VOP_UNLOCK(vp, 0);
139 	if (error)
140 		goto fail;
141 	audit_fstat.af_currsz = vattr.va_size;
142 
143 	/*
144 	 * We handle four different space-related limits:
145 	 *
146 	 * - A fixed (hard) limit on the minimum free blocks we require on
147 	 *   the file system, and results in record loss, a trigger, and
148 	 *   possible fail stop due to violating invariants.
149 	 *
150 	 * - An administrative (soft) limit, which when fallen below, results
151 	 *   in the kernel notifying the audit daemon of low space.
152 	 *
153 	 * - An audit trail size limit, which when gone above, results in the
154 	 *   kernel notifying the audit daemon that rotation is desired.
155 	 *
156 	 * - The total depth of the kernel audit record exceeding free space,
157 	 *   which can lead to possible fail stop (with drain), in order to
158 	 *   prevent violating invariants.  Failure here doesn't halt
159 	 *   immediately, but prevents new records from being generated.
160 	 *
161 	 * Possibly, the last of these should be handled differently, always
162 	 * allowing a full queue to be lost, rather than trying to prevent
163 	 * loss.
164 	 *
165 	 * First, handle the hard limit, which generates a trigger and may
166 	 * fail stop.  This is handled in the same manner as ENOSPC from
167 	 * VOP_WRITE, and results in record loss.
168 	 */
169 	if (mnt_stat->f_bfree < AUDIT_HARD_LIMIT_FREE_BLOCKS) {
170 		error = ENOSPC;
171 		goto fail_enospc;
172 	}
173 
174 	/*
175 	 * Second, handle falling below the soft limit, if defined; we send
176 	 * the daemon a trigger and continue processing the record.  Triggers
177 	 * are limited to 1/sec.
178 	 */
179 	if (audit_qctrl.aq_minfree != 0) {
180 		temp = mnt_stat->f_blocks / (100 / audit_qctrl.aq_minfree);
181 		if (mnt_stat->f_bfree < temp) {
182 			if (ppsratecheck(&last_lowspace_trigger,
183 			    &cur_lowspace_trigger, 1)) {
184 				(void)audit_send_trigger(
185 				    AUDIT_TRIGGER_LOW_SPACE);
186 				printf("Warning: disk space low (< %d%% free) "
187 				    "on audit log file-system\n",
188 				    audit_qctrl.aq_minfree);
189 			}
190 		}
191 	}
192 
193 	/*
194 	 * If the current file is getting full, generate a rotation trigger
195 	 * to the daemon.  This is only approximate, which is fine as more
196 	 * records may be generated before the daemon rotates the file.
197 	 */
198 	if ((audit_fstat.af_filesz != 0) && (audit_file_rotate_wait == 0) &&
199 	    (vattr.va_size >= audit_fstat.af_filesz)) {
200 		AUDIT_WORKER_LOCK_ASSERT();
201 
202 		audit_file_rotate_wait = 1;
203 		(void)audit_send_trigger(AUDIT_TRIGGER_ROTATE_KERNEL);
204 	}
205 
206 	/*
207 	 * If the estimated amount of audit data in the audit event queue
208 	 * (plus records allocated but not yet queued) has reached the amount
209 	 * of free space on the disk, then we need to go into an audit fail
210 	 * stop state, in which we do not permit the allocation/committing of
211 	 * any new audit records.  We continue to process records but don't
212 	 * allow any activities that might generate new records.  In the
213 	 * future, we might want to detect when space is available again and
214 	 * allow operation to continue, but this behavior is sufficient to
215 	 * meet fail stop requirements in CAPP.
216 	 */
217 	if (audit_fail_stop) {
218 		if ((unsigned long)((audit_q_len + audit_pre_q_len + 1) *
219 		    MAX_AUDIT_RECORD_SIZE) / mnt_stat->f_bsize >=
220 		    (unsigned long)(mnt_stat->f_bfree)) {
221 			if (ppsratecheck(&last_fail, &cur_fail, 1))
222 				printf("audit_record_write: free space "
223 				    "below size of audit queue, failing "
224 				    "stop\n");
225 			audit_in_failure = 1;
226 		} else if (audit_in_failure) {
227 			/*
228 			 * Note: if we want to handle recovery, this is the
229 			 * spot to do it: unset audit_in_failure, and issue a
230 			 * wakeup on the cv.
231 			 */
232 		}
233 	}
234 
235 	error = vn_rdwr(UIO_WRITE, vp, data, len, (off_t)0, UIO_SYSSPACE,
236 	    IO_APPEND|IO_UNIT, cred, NULL, NULL, curthread);
237 	if (error == ENOSPC)
238 		goto fail_enospc;
239 	else if (error)
240 		goto fail;
241 
242 	/*
243 	 * Catch completion of a queue drain here; if we're draining and the
244 	 * queue is now empty, fail stop.  That audit_fail_stop is implicitly
245 	 * true, since audit_in_failure can only be set of audit_fail_stop is
246 	 * set.
247 	 *
248 	 * Note: if we handle recovery from audit_in_failure, then we need to
249 	 * make panic here conditional.
250 	 */
251 	if (audit_in_failure) {
252 		if (audit_q_len == 0 && audit_pre_q_len == 0) {
253 			VOP_LOCK(vp, LK_EXCLUSIVE | LK_RETRY);
254 			(void)VOP_FSYNC(vp, MNT_WAIT, curthread);
255 			VOP_UNLOCK(vp, 0);
256 			panic("Audit store overflow; record queue drained.");
257 		}
258 	}
259 
260 	return;
261 
262 fail_enospc:
263 	/*
264 	 * ENOSPC is considered a special case with respect to failures, as
265 	 * this can reflect either our preemptive detection of insufficient
266 	 * space, or ENOSPC returned by the vnode write call.
267 	 */
268 	if (audit_fail_stop) {
269 		VOP_LOCK(vp, LK_EXCLUSIVE | LK_RETRY);
270 		(void)VOP_FSYNC(vp, MNT_WAIT, curthread);
271 		VOP_UNLOCK(vp, 0);
272 		panic("Audit log space exhausted and fail-stop set.");
273 	}
274 	(void)audit_send_trigger(AUDIT_TRIGGER_NO_SPACE);
275 	audit_suspended = 1;
276 
277 	/* FALLTHROUGH */
278 fail:
279 	/*
280 	 * We have failed to write to the file, so the current record is
281 	 * lost, which may require an immediate system halt.
282 	 */
283 	if (audit_panic_on_write_fail) {
284 		VOP_LOCK(vp, LK_EXCLUSIVE | LK_RETRY);
285 		(void)VOP_FSYNC(vp, MNT_WAIT, curthread);
286 		VOP_UNLOCK(vp, 0);
287 		panic("audit_worker: write error %d\n", error);
288 	} else if (ppsratecheck(&last_fail, &cur_fail, 1))
289 		printf("audit_worker: write error %d\n", error);
290 }
291 
292 /*
293  * Given a kernel audit record, process as required.  Kernel audit records
294  * are converted to one, or possibly two, BSM records, depending on whether
295  * there is a user audit record present also.  Kernel records need be
296  * converted to BSM before they can be written out.  Both types will be
297  * written to disk, and audit pipes.
298  */
299 static void
300 audit_worker_process_record(struct kaudit_record *ar)
301 {
302 	struct au_record *bsm;
303 	au_class_t class;
304 	au_event_t event;
305 	au_id_t auid;
306 	int error, sorf;
307 	int locked;
308 
309 	/*
310 	 * We hold the audit worker lock over both writes, if there are two,
311 	 * so that the two records won't be split across a rotation and end
312 	 * up in two different trail files.
313 	 */
314 	if (((ar->k_ar_commit & AR_COMMIT_USER) &&
315 	    (ar->k_ar_commit & AR_PRESELECT_USER_TRAIL)) ||
316 	    (ar->k_ar_commit & AR_PRESELECT_TRAIL)) {
317 		AUDIT_WORKER_LOCK();
318 		locked = 1;
319 	} else
320 		locked = 0;
321 
322 	/*
323 	 * First, handle the user record, if any: commit to the system trail
324 	 * and audit pipes as selected.
325 	 */
326 	if ((ar->k_ar_commit & AR_COMMIT_USER) &&
327 	    (ar->k_ar_commit & AR_PRESELECT_USER_TRAIL)) {
328 		AUDIT_WORKER_LOCK_ASSERT();
329 		audit_record_write(audit_vp, audit_cred, ar->k_udata,
330 		    ar->k_ulen);
331 	}
332 
333 	if ((ar->k_ar_commit & AR_COMMIT_USER) &&
334 	    (ar->k_ar_commit & AR_PRESELECT_USER_PIPE))
335 		audit_pipe_submit_user(ar->k_udata, ar->k_ulen);
336 
337 	if (!(ar->k_ar_commit & AR_COMMIT_KERNEL) ||
338 	    ((ar->k_ar_commit & AR_PRESELECT_PIPE) == 0 &&
339 	    (ar->k_ar_commit & AR_PRESELECT_TRAIL) == 0))
340 		goto out;
341 
342 	auid = ar->k_ar.ar_subj_auid;
343 	event = ar->k_ar.ar_event;
344 	class = au_event_class(event);
345 	if (ar->k_ar.ar_errno == 0)
346 		sorf = AU_PRS_SUCCESS;
347 	else
348 		sorf = AU_PRS_FAILURE;
349 
350 	error = kaudit_to_bsm(ar, &bsm);
351 	switch (error) {
352 	case BSM_NOAUDIT:
353 		goto out;
354 
355 	case BSM_FAILURE:
356 		printf("audit_worker_process_record: BSM_FAILURE\n");
357 		goto out;
358 
359 	case BSM_SUCCESS:
360 		break;
361 
362 	default:
363 		panic("kaudit_to_bsm returned %d", error);
364 	}
365 
366 	if (ar->k_ar_commit & AR_PRESELECT_TRAIL) {
367 		AUDIT_WORKER_LOCK_ASSERT();
368 		audit_record_write(audit_vp, audit_cred, bsm->data, bsm->len);
369 	}
370 
371 	if (ar->k_ar_commit & AR_PRESELECT_PIPE)
372 		audit_pipe_submit(auid, event, class, sorf,
373 		    ar->k_ar_commit & AR_PRESELECT_TRAIL, bsm->data,
374 		    bsm->len);
375 
376 	kau_free(bsm);
377 out:
378 	if (locked)
379 		AUDIT_WORKER_UNLOCK();
380 }
381 
382 /*
383  * The audit_worker thread is responsible for watching the event queue,
384  * dequeueing records, converting them to BSM format, and committing them to
385  * disk.  In order to minimize lock thrashing, records are dequeued in sets
386  * to a thread-local work queue.
387  *
388  * Note: this means that the effect bound on the size of the pending record
389  * queue is 2x the length of the global queue.
390  */
391 static void
392 audit_worker(void *arg)
393 {
394 	struct kaudit_queue ar_worklist;
395 	struct kaudit_record *ar;
396 	int lowater_signal;
397 
398 	TAILQ_INIT(&ar_worklist);
399 	mtx_lock(&audit_mtx);
400 	while (1) {
401 		mtx_assert(&audit_mtx, MA_OWNED);
402 
403 		/*
404 		 * Wait for a record.
405 		 */
406 		while (TAILQ_EMPTY(&audit_q))
407 			cv_wait(&audit_worker_cv, &audit_mtx);
408 
409 		/*
410 		 * If there are records in the global audit record queue,
411 		 * transfer them to a thread-local queue and process them
412 		 * one by one.  If we cross the low watermark threshold,
413 		 * signal any waiting processes that they may wake up and
414 		 * continue generating records.
415 		 */
416 		lowater_signal = 0;
417 		while ((ar = TAILQ_FIRST(&audit_q))) {
418 			TAILQ_REMOVE(&audit_q, ar, k_q);
419 			audit_q_len--;
420 			if (audit_q_len == audit_qctrl.aq_lowater)
421 				lowater_signal++;
422 			TAILQ_INSERT_TAIL(&ar_worklist, ar, k_q);
423 		}
424 		if (lowater_signal)
425 			cv_broadcast(&audit_watermark_cv);
426 
427 		mtx_unlock(&audit_mtx);
428 		while ((ar = TAILQ_FIRST(&ar_worklist))) {
429 			TAILQ_REMOVE(&ar_worklist, ar, k_q);
430 			audit_worker_process_record(ar);
431 			audit_free(ar);
432 		}
433 		mtx_lock(&audit_mtx);
434 	}
435 }
436 
437 /*
438  * audit_rotate_vnode() is called by a user or kernel thread to configure or
439  * de-configure auditing on a vnode.  The arguments are the replacement
440  * credential (referenced) and vnode (referenced and opened) to substitute
441  * for the current credential and vnode, if any.  If either is set to NULL,
442  * both should be NULL, and this is used to indicate that audit is being
443  * disabled.  Any previous cred/vnode will be closed and freed.  We re-enable
444  * generating rotation requests to auditd.
445  */
446 void
447 audit_rotate_vnode(struct ucred *cred, struct vnode *vp)
448 {
449 	struct ucred *old_audit_cred;
450 	struct vnode *old_audit_vp;
451 
452 	KASSERT((cred != NULL && vp != NULL) || (cred == NULL && vp == NULL),
453 	    ("audit_rotate_vnode: cred %p vp %p", cred, vp));
454 
455 	/*
456 	 * Rotate the vnode/cred, and clear the rotate flag so that we will
457 	 * send a rotate trigger if the new file fills.
458 	 */
459 	AUDIT_WORKER_LOCK();
460 	old_audit_cred = audit_cred;
461 	old_audit_vp = audit_vp;
462 	audit_cred = cred;
463 	audit_vp = vp;
464 	audit_file_rotate_wait = 0;
465 	audit_enabled = (audit_vp != NULL);
466 	AUDIT_WORKER_UNLOCK();
467 
468 	/*
469 	 * If there was an old vnode/credential, close and free.
470 	 */
471 	if (old_audit_vp != NULL) {
472 		vn_close(old_audit_vp, AUDIT_CLOSE_FLAGS, old_audit_cred,
473 		    curthread);
474 		crfree(old_audit_cred);
475 	}
476 }
477 
478 void
479 audit_worker_init(void)
480 {
481 	int error;
482 
483 	AUDIT_WORKER_LOCK_INIT();
484 	error = kproc_create(audit_worker, NULL, &audit_thread, RFHIGHPID,
485 	    0, "audit");
486 	if (error)
487 		panic("audit_worker_init: kproc_create returned %d", error);
488 }
489