xref: /freebsd/sys/security/audit/audit_worker.c (revision 10f0bcab61ef441cb5af32fb706688d8cbd55dc0)
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  * $FreeBSD$
31  */
32 
33 #include <sys/param.h>
34 #include <sys/condvar.h>
35 #include <sys/conf.h>
36 #include <sys/file.h>
37 #include <sys/filedesc.h>
38 #include <sys/fcntl.h>
39 #include <sys/ipc.h>
40 #include <sys/kernel.h>
41 #include <sys/kthread.h>
42 #include <sys/malloc.h>
43 #include <sys/mount.h>
44 #include <sys/namei.h>
45 #include <sys/proc.h>
46 #include <sys/queue.h>
47 #include <sys/socket.h>
48 #include <sys/socketvar.h>
49 #include <sys/protosw.h>
50 #include <sys/domain.h>
51 #include <sys/sx.h>
52 #include <sys/sysproto.h>
53 #include <sys/sysent.h>
54 #include <sys/systm.h>
55 #include <sys/ucred.h>
56 #include <sys/uio.h>
57 #include <sys/un.h>
58 #include <sys/unistd.h>
59 #include <sys/vnode.h>
60 
61 #include <bsm/audit.h>
62 #include <bsm/audit_internal.h>
63 #include <bsm/audit_kevents.h>
64 
65 #include <netinet/in.h>
66 #include <netinet/in_pcb.h>
67 
68 #include <security/audit/audit.h>
69 #include <security/audit/audit_private.h>
70 
71 #include <vm/uma.h>
72 
73 /*
74  * Worker thread that will schedule disk I/O, etc.
75  */
76 static struct proc		*audit_thread;
77 
78 /*
79  * audit_cred and audit_vp are the stored credential and vnode to use for
80  * active audit trail.  They are protected by audit_worker_sx, which will be
81  * held across all I/O and all rotation to prevent them from being replaced
82  * (rotated) while in use.  The audit_file_rotate_wait flag is set when the
83  * kernel has delivered a trigger to auditd to rotate the trail, and is
84  * cleared when the next rotation takes place.  It is also protected by
85  * audit_worker_sx.
86  */
87 static int		 audit_file_rotate_wait;
88 static struct sx	 audit_worker_sx;
89 static struct ucred	*audit_cred;
90 static struct vnode	*audit_vp;
91 
92 /*
93  * Write an audit record to a file, performed as the last stage after both
94  * preselection and BSM conversion.  Both space management and write failures
95  * are handled in this function.
96  *
97  * No attempt is made to deal with possible failure to deliver a trigger to
98  * the audit daemon, since the message is asynchronous anyway.
99  */
100 static void
101 audit_record_write(struct vnode *vp, struct ucred *cred, void *data,
102     size_t len)
103 {
104 	static struct timeval last_lowspace_trigger;
105 	static struct timeval last_fail;
106 	static int cur_lowspace_trigger;
107 	struct statfs *mnt_stat;
108 	int error, vfslocked;
109 	static int cur_fail;
110 	struct vattr vattr;
111 	long temp;
112 
113 	sx_assert(&audit_worker_sx, SA_LOCKED);	/* audit_file_rotate_wait. */
114 
115 	if (vp == NULL)
116 		return;
117 
118  	mnt_stat = &vp->v_mount->mnt_stat;
119 	vfslocked = VFS_LOCK_GIANT(vp->v_mount);
120 
121 	/*
122 	 * First, gather statistics on the audit log file and file system so
123 	 * that we know how we're doing on space.  Consider failure of these
124 	 * operations to indicate a future inability to write to the file.
125 	 */
126 	error = VFS_STATFS(vp->v_mount, mnt_stat, curthread);
127 	if (error)
128 		goto fail;
129 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
130 	error = VOP_GETATTR(vp, &vattr, cred, curthread);
131 	VOP_UNLOCK(vp, 0);
132 	if (error)
133 		goto fail;
134 	audit_fstat.af_currsz = vattr.va_size;
135 
136 	/*
137 	 * We handle four different space-related limits:
138 	 *
139 	 * - A fixed (hard) limit on the minimum free blocks we require on
140 	 *   the file system, and results in record loss, a trigger, and
141 	 *   possible fail stop due to violating invariants.
142 	 *
143 	 * - An administrative (soft) limit, which when fallen below, results
144 	 *   in the kernel notifying the audit daemon of low space.
145 	 *
146 	 * - An audit trail size limit, which when gone above, results in the
147 	 *   kernel notifying the audit daemon that rotation is desired.
148 	 *
149 	 * - The total depth of the kernel audit record exceeding free space,
150 	 *   which can lead to possible fail stop (with drain), in order to
151 	 *   prevent violating invariants.  Failure here doesn't halt
152 	 *   immediately, but prevents new records from being generated.
153 	 *
154 	 * Possibly, the last of these should be handled differently, always
155 	 * allowing a full queue to be lost, rather than trying to prevent
156 	 * loss.
157 	 *
158 	 * First, handle the hard limit, which generates a trigger and may
159 	 * fail stop.  This is handled in the same manner as ENOSPC from
160 	 * VOP_WRITE, and results in record loss.
161 	 */
162 	if (mnt_stat->f_bfree < AUDIT_HARD_LIMIT_FREE_BLOCKS) {
163 		error = ENOSPC;
164 		goto fail_enospc;
165 	}
166 
167 	/*
168 	 * Second, handle falling below the soft limit, if defined; we send
169 	 * the daemon a trigger and continue processing the record.  Triggers
170 	 * are limited to 1/sec.
171 	 */
172 	if (audit_qctrl.aq_minfree != 0) {
173 		temp = mnt_stat->f_blocks / (100 / audit_qctrl.aq_minfree);
174 		if (mnt_stat->f_bfree < temp) {
175 			if (ppsratecheck(&last_lowspace_trigger,
176 			    &cur_lowspace_trigger, 1)) {
177 				(void)audit_send_trigger(
178 				    AUDIT_TRIGGER_LOW_SPACE);
179 				printf("Warning: audit space low\n");
180 			}
181 		}
182 	}
183 
184 	/*
185 	 * If the current file is getting full, generate a rotation trigger
186 	 * to the daemon.  This is only approximate, which is fine as more
187 	 * records may be generated before the daemon rotates the file.
188 	 */
189 	if ((audit_fstat.af_filesz != 0) && (audit_file_rotate_wait == 0) &&
190 	    (vattr.va_size >= audit_fstat.af_filesz)) {
191 		sx_assert(&audit_worker_sx, SA_XLOCKED);
192 
193 		audit_file_rotate_wait = 1;
194 		(void)audit_send_trigger(AUDIT_TRIGGER_ROTATE_KERNEL);
195 	}
196 
197 	/*
198 	 * If the estimated amount of audit data in the audit event queue
199 	 * (plus records allocated but not yet queued) has reached the amount
200 	 * of free space on the disk, then we need to go into an audit fail
201 	 * stop state, in which we do not permit the allocation/committing of
202 	 * any new audit records.  We continue to process records but don't
203 	 * allow any activities that might generate new records.  In the
204 	 * future, we might want to detect when space is available again and
205 	 * allow operation to continue, but this behavior is sufficient to
206 	 * meet fail stop requirements in CAPP.
207 	 */
208 	if (audit_fail_stop) {
209 		if ((unsigned long)((audit_q_len + audit_pre_q_len + 1) *
210 		    MAX_AUDIT_RECORD_SIZE) / mnt_stat->f_bsize >=
211 		    (unsigned long)(mnt_stat->f_bfree)) {
212 			if (ppsratecheck(&last_fail, &cur_fail, 1))
213 				printf("audit_record_write: free space "
214 				    "below size of audit queue, failing "
215 				    "stop\n");
216 			audit_in_failure = 1;
217 		} else if (audit_in_failure) {
218 			/*
219 			 * Note: if we want to handle recovery, this is the
220 			 * spot to do it: unset audit_in_failure, and issue a
221 			 * wakeup on the cv.
222 			 */
223 		}
224 	}
225 
226 	error = vn_rdwr(UIO_WRITE, vp, data, len, (off_t)0, UIO_SYSSPACE,
227 	    IO_APPEND|IO_UNIT, cred, NULL, NULL, curthread);
228 	if (error == ENOSPC)
229 		goto fail_enospc;
230 	else if (error)
231 		goto fail;
232 
233 	/*
234 	 * Catch completion of a queue drain here; if we're draining and the
235 	 * queue is now empty, fail stop.  That audit_fail_stop is implicitly
236 	 * true, since audit_in_failure can only be set of audit_fail_stop is
237 	 * set.
238 	 *
239 	 * Note: if we handle recovery from audit_in_failure, then we need to
240 	 * make panic here conditional.
241 	 */
242 	if (audit_in_failure) {
243 		if (audit_q_len == 0 && audit_pre_q_len == 0) {
244 			VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK);
245 			(void)VOP_FSYNC(vp, MNT_WAIT, curthread);
246 			VOP_UNLOCK(vp, 0);
247 			panic("Audit store overflow; record queue drained.");
248 		}
249 	}
250 
251 	VFS_UNLOCK_GIANT(vfslocked);
252 	return;
253 
254 fail_enospc:
255 	/*
256 	 * ENOSPC is considered a special case with respect to failures, as
257 	 * this can reflect either our preemptive detection of insufficient
258 	 * space, or ENOSPC returned by the vnode write call.
259 	 */
260 	if (audit_fail_stop) {
261 		VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK);
262 		(void)VOP_FSYNC(vp, MNT_WAIT, curthread);
263 		VOP_UNLOCK(vp, 0);
264 		panic("Audit log space exhausted and fail-stop set.");
265 	}
266 	(void)audit_send_trigger(AUDIT_TRIGGER_NO_SPACE);
267 	audit_suspended = 1;
268 
269 	/* FALLTHROUGH */
270 fail:
271 	/*
272 	 * We have failed to write to the file, so the current record is
273 	 * lost, which may require an immediate system halt.
274 	 */
275 	if (audit_panic_on_write_fail) {
276 		VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK);
277 		(void)VOP_FSYNC(vp, MNT_WAIT, curthread);
278 		VOP_UNLOCK(vp, 0);
279 		panic("audit_worker: write error %d\n", error);
280 	} else if (ppsratecheck(&last_fail, &cur_fail, 1))
281 		printf("audit_worker: write error %d\n", error);
282 	VFS_UNLOCK_GIANT(vfslocked);
283 }
284 
285 /*
286  * Given a kernel audit record, process as required.  Kernel audit records
287  * are converted to one, or possibly two, BSM records, depending on whether
288  * there is a user audit record present also.  Kernel records need be
289  * converted to BSM before they can be written out.  Both types will be
290  * written to disk, and audit pipes.
291  */
292 static void
293 audit_worker_process_record(struct kaudit_record *ar)
294 {
295 	struct au_record *bsm;
296 	au_class_t class;
297 	au_event_t event;
298 	au_id_t auid;
299 	int error, sorf;
300 	int trail_locked;
301 
302 	/*
303 	 * We hold the audit_worker_sx lock over both writes, if there are
304 	 * two, so that the two records won't be split across a rotation and
305 	 * end up in two different trail files.
306 	 */
307 	if (((ar->k_ar_commit & AR_COMMIT_USER) &&
308 	    (ar->k_ar_commit & AR_PRESELECT_USER_TRAIL)) ||
309 	    (ar->k_ar_commit & AR_PRESELECT_TRAIL)) {
310 		sx_xlock(&audit_worker_sx);
311 		trail_locked = 1;
312 	} else
313 		trail_locked = 0;
314 
315 	/*
316 	 * First, handle the user record, if any: commit to the system trail
317 	 * and audit pipes as selected.
318 	 */
319 	if ((ar->k_ar_commit & AR_COMMIT_USER) &&
320 	    (ar->k_ar_commit & AR_PRESELECT_USER_TRAIL)) {
321 		sx_assert(&audit_worker_sx, SA_XLOCKED);
322 		audit_record_write(audit_vp, audit_cred, ar->k_udata,
323 		    ar->k_ulen);
324 	}
325 
326 	if ((ar->k_ar_commit & AR_COMMIT_USER) &&
327 	    (ar->k_ar_commit & AR_PRESELECT_USER_PIPE))
328 		audit_pipe_submit_user(ar->k_udata, ar->k_ulen);
329 
330 	if (!(ar->k_ar_commit & AR_COMMIT_KERNEL) ||
331 	    ((ar->k_ar_commit & AR_PRESELECT_PIPE) == 0 &&
332 	    (ar->k_ar_commit & AR_PRESELECT_TRAIL) == 0))
333 		goto out;
334 
335 	auid = ar->k_ar.ar_subj_auid;
336 	event = ar->k_ar.ar_event;
337 	class = au_event_class(event);
338 	if (ar->k_ar.ar_errno == 0)
339 		sorf = AU_PRS_SUCCESS;
340 	else
341 		sorf = AU_PRS_FAILURE;
342 
343 	error = kaudit_to_bsm(ar, &bsm);
344 	switch (error) {
345 	case BSM_NOAUDIT:
346 		goto out;
347 
348 	case BSM_FAILURE:
349 		printf("audit_worker_process_record: BSM_FAILURE\n");
350 		goto out;
351 
352 	case BSM_SUCCESS:
353 		break;
354 
355 	default:
356 		panic("kaudit_to_bsm returned %d", error);
357 	}
358 
359 	if (ar->k_ar_commit & AR_PRESELECT_TRAIL) {
360 		sx_assert(&audit_worker_sx, SA_XLOCKED);
361 		audit_record_write(audit_vp, audit_cred, bsm->data, bsm->len);
362 	}
363 
364 	if (ar->k_ar_commit & AR_PRESELECT_PIPE)
365 		audit_pipe_submit(auid, event, class, sorf,
366 		    ar->k_ar_commit & AR_PRESELECT_TRAIL, bsm->data,
367 		    bsm->len);
368 
369 	kau_free(bsm);
370 out:
371 	if (trail_locked)
372 		sx_xunlock(&audit_worker_sx);
373 }
374 
375 /*
376  * The audit_worker thread is responsible for watching the event queue,
377  * dequeueing records, converting them to BSM format, and committing them to
378  * disk.  In order to minimize lock thrashing, records are dequeued in sets
379  * to a thread-local work queue.
380  *
381  * Note: this means that the effect bound on the size of the pending record
382  * queue is 2x the length of the global queue.
383  */
384 static void
385 audit_worker(void *arg)
386 {
387 	struct kaudit_queue ar_worklist;
388 	struct kaudit_record *ar;
389 	int lowater_signal;
390 
391 	TAILQ_INIT(&ar_worklist);
392 	mtx_lock(&audit_mtx);
393 	while (1) {
394 		mtx_assert(&audit_mtx, MA_OWNED);
395 
396 		/*
397 		 * Wait for a record.
398 		 */
399 		while (TAILQ_EMPTY(&audit_q))
400 			cv_wait(&audit_worker_cv, &audit_mtx);
401 
402 		/*
403 		 * If there are records in the global audit record queue,
404 		 * transfer them to a thread-local queue and process them
405 		 * one by one.  If we cross the low watermark threshold,
406 		 * signal any waiting processes that they may wake up and
407 		 * continue generating records.
408 		 */
409 		lowater_signal = 0;
410 		while ((ar = TAILQ_FIRST(&audit_q))) {
411 			TAILQ_REMOVE(&audit_q, ar, k_q);
412 			audit_q_len--;
413 			if (audit_q_len == audit_qctrl.aq_lowater)
414 				lowater_signal++;
415 			TAILQ_INSERT_TAIL(&ar_worklist, ar, k_q);
416 		}
417 		if (lowater_signal)
418 			cv_broadcast(&audit_watermark_cv);
419 
420 		mtx_unlock(&audit_mtx);
421 		while ((ar = TAILQ_FIRST(&ar_worklist))) {
422 			TAILQ_REMOVE(&ar_worklist, ar, k_q);
423 			audit_worker_process_record(ar);
424 			audit_free(ar);
425 		}
426 		mtx_lock(&audit_mtx);
427 	}
428 }
429 
430 /*
431  * audit_rotate_vnode() is called by a user or kernel thread to configure or
432  * de-configure auditing on a vnode.  The arguments are the replacement
433  * credential (referenced) and vnode (referenced and opened) to substitute
434  * for the current credential and vnode, if any.  If either is set to NULL,
435  * both should be NULL, and this is used to indicate that audit is being
436  * disabled.  Any previous cred/vnode will be closed and freed.  We re-enable
437  * generating rotation requests to auditd.
438  */
439 void
440 audit_rotate_vnode(struct ucred *cred, struct vnode *vp)
441 {
442 	struct ucred *old_audit_cred;
443 	struct vnode *old_audit_vp;
444 	int vfslocked;
445 
446 	KASSERT((cred != NULL && vp != NULL) || (cred == NULL && vp == NULL),
447 	    ("audit_rotate_vnode: cred %p vp %p", cred, vp));
448 
449 	/*
450 	 * Rotate the vnode/cred, and clear the rotate flag so that we will
451 	 * send a rotate trigger if the new file fills.
452 	 */
453 	sx_xlock(&audit_worker_sx);
454 	old_audit_cred = audit_cred;
455 	old_audit_vp = audit_vp;
456 	audit_cred = cred;
457 	audit_vp = vp;
458 	audit_file_rotate_wait = 0;
459 	audit_enabled = (audit_vp != NULL);
460 	sx_xunlock(&audit_worker_sx);
461 
462 	/*
463 	 * If there was an old vnode/credential, close and free.
464 	 */
465 	if (old_audit_vp != NULL) {
466 		vfslocked = VFS_LOCK_GIANT(old_audit_vp->v_mount);
467 		vn_close(old_audit_vp, AUDIT_CLOSE_FLAGS, old_audit_cred,
468 		    curthread);
469 		VFS_UNLOCK_GIANT(vfslocked);
470 		crfree(old_audit_cred);
471 	}
472 }
473 
474 void
475 audit_worker_init(void)
476 {
477 	int error;
478 
479 	sx_init(&audit_worker_sx, "audit_worker_sx");
480 	error = kproc_create(audit_worker, NULL, &audit_thread, RFHIGHPID,
481 	    0, "audit");
482 	if (error)
483 		panic("audit_worker_init: kproc_create returned %d", error);
484 }
485