1/*-
2 * Copyright (c) 1999-2016 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/param.h>
32#include <sys/fcntl.h>
33#include <sys/kernel.h>
34#include <sys/lock.h>
35#include <sys/namei.h>
36#include <sys/proc_internal.h>
37#include <sys/kauth.h>
38#include <sys/queue.h>
39#include <sys/systm.h>
40#include <sys/time.h>
41#include <sys/ucred.h>
42#include <sys/uio.h>
43#include <sys/unistd.h>
44#include <sys/file_internal.h>
45#include <sys/vnode_internal.h>
46#include <sys/user.h>
47#include <sys/syscall.h>
48#include <sys/malloc.h>
49#include <sys/un.h>
50#include <sys/sysent.h>
51#include <sys/sysproto.h>
52#include <sys/vfs_context.h>
53#include <sys/domain.h>
54#include <sys/protosw.h>
55#include <sys/socketvar.h>
56
57#include <bsm/audit.h>
58#include <bsm/audit_internal.h>
59#include <bsm/audit_kevents.h>
60
61#include <security/audit/audit.h>
62#include <security/audit/audit_bsd.h>
63#include <security/audit/audit_private.h>
64
65#include <mach/host_priv.h>
66#include <mach/host_special_ports.h>
67#include <mach/audit_triggers_server.h>
68
69#include <kern/host.h>
70#include <kern/zalloc.h>
71#include <kern/sched_prim.h>
72#include <kern/task.h>
73
74#include <net/route.h>
75
76#include <netinet/in.h>
77#include <netinet/in_pcb.h>
78
79/*
80 * Worker thread that will schedule disk I/O, etc.
81 */
82static thread_t audit_thread;
83
84/*
85 * audit_ctx and audit_vp are the stored credential and vnode to use for
86 * active audit trail. They are protected by audit_worker_sl, which will be
87 * held across all I/O and all rotation to prevent them from being replaced
88 * (rotated) while in use. The audit_file_rotate_wait flag is set when the
89 * kernel has delivered a trigger to auditd to rotate the trail, and is
90 * cleared when the next rotation takes place. It is also protected by
91 * audit_worker_sl.
92 */
93static int audit_file_rotate_wait;
94static struct slck audit_worker_sl;
95static struct vfs_context audit_ctx;
96static struct vnode *audit_vp;
97
98#define AUDIT_WORKER_SX_INIT() slck_init(&audit_worker_sl, \
99 "audit_worker_sl")
100#define AUDIT_WORKER_SX_XLOCK() slck_lock(&audit_worker_sl)
101#define AUDIT_WORKER_SX_XUNLOCK() slck_unlock(&audit_worker_sl)
102#define AUDIT_WORKER_SX_ASSERT() slck_assert(&audit_worker_sl, SL_OWNED)
103#define AUDIT_WORKER_SX_DESTROY() slck_destroy(&audit_worker_sl)
104
105/*
106 * The audit_q_draining flag is set when audit is disabled and the audit
107 * worker queue is being drained.
108 */
109static int audit_q_draining;
110
111/*
112 * The special kernel audit record, audit_drain_kar, is used to mark the end of
113 * the queue when draining it.
114 */
115static struct kaudit_record audit_drain_kar = {
116 .k_ar = {
117 .ar_event = AUE_NULL,
118 },
119 .k_ar_commit = AR_DRAIN_QUEUE,
120};
121
122/*
123 * Write an audit record to a file, performed as the last stage after both
124 * preselection and BSM conversion. Both space management and write failures
125 * are handled in this function.
126 *
127 * No attempt is made to deal with possible failure to deliver a trigger to
128 * the audit daemon, since the message is asynchronous anyway.
129 */
130static void
131audit_record_write(struct vnode *vp, struct vfs_context *ctx, void *data,
132 size_t len)
133{
134 static struct timeval last_lowspace_trigger;
135 static struct timeval last_fail;
136 static int cur_lowspace_trigger;
137 struct vfsstatfs *mnt_stat;
138 int error;
139 static int cur_fail;
140 uint64_t temp;
141 off_t file_size;
142
143 AUDIT_WORKER_SX_ASSERT(); /* audit_file_rotate_wait. */
144
145 if (vp == NULL)
146 return;
147
148 if (vnode_getwithref(vp))
149 return /*(ENOENT)*/;
150
151 mnt_stat = &vp->v_mount->mnt_vfsstat;
152
153 /*
154 * First, gather statistics on the audit log file and file system so
155 * that we know how we're doing on space. Consider failure of these
156 * operations to indicate a future inability to write to the file.
157 */
158 error = vfs_update_vfsstat(vp->v_mount, ctx, VFS_KERNEL_EVENT);
159 if (error)
160 goto fail;
161 error = vnode_size(vp, &file_size, ctx);
162 if (error)
163 goto fail;
164 audit_fstat.af_currsz = (u_quad_t)file_size;
165
166 /*
167 * We handle four different space-related limits:
168 *
169 * - A fixed (hard) limit on the minimum free blocks we require on
170 * the file system, and results in record loss, a trigger, and
171 * possible fail stop due to violating invariants.
172 *
173 * - An administrative (soft) limit, which when fallen below, results
174 * in the kernel notifying the audit daemon of low space.
175 *
176 * - An audit trail size limit, which when gone above, results in the
177 * kernel notifying the audit daemon that rotation is desired.
178 *
179 * - The total depth of the kernel audit record exceeding free space,
180 * which can lead to possible fail stop (with drain), in order to
181 * prevent violating invariants. Failure here doesn't halt
182 * immediately, but prevents new records from being generated.
183 *
184 * Possibly, the last of these should be handled differently, always
185 * allowing a full queue to be lost, rather than trying to prevent
186 * loss.
187 *
188 * First, handle the hard limit, which generates a trigger and may
189 * fail stop. This is handled in the same manner as ENOSPC from
190 * VOP_WRITE, and results in record loss.
191 */
192 if (mnt_stat->f_bfree < AUDIT_HARD_LIMIT_FREE_BLOCKS) {
193 error = ENOSPC;
194 goto fail_enospc;
195 }
196
197 /*
198 * Second, handle falling below the soft limit, if defined; we send
199 * the daemon a trigger and continue processing the record. Triggers
200 * are limited to 1/sec.
201 */
202 if (audit_qctrl.aq_minfree != 0) {
203 temp = mnt_stat->f_blocks / (100 / audit_qctrl.aq_minfree);
204 if (mnt_stat->f_bfree < temp &&
205 ppsratecheck(&last_lowspace_trigger,
206 &cur_lowspace_trigger, 1))
207 (void)audit_send_trigger(
208 AUDIT_TRIGGER_LOW_SPACE);
209 }
210
211 /*
212 * If the current file is getting full, generate a rotation trigger
213 * to the daemon. This is only approximate, which is fine as more
214 * records may be generated before the daemon rotates the file.
215 */
216 if ((audit_fstat.af_filesz != 0) && (audit_file_rotate_wait == 0) &&
217 ((u_quad_t)file_size >= audit_fstat.af_filesz)) {
218 AUDIT_WORKER_SX_ASSERT();
219
220 audit_file_rotate_wait = 1;
221 (void)audit_send_trigger(AUDIT_TRIGGER_ROTATE_KERNEL);
222 }
223
224 /*
225 * If the estimated amount of audit data in the audit event queue
226 * (plus records allocated but not yet queued) has reached the amount
227 * of free space on the disk, then we need to go into an audit fail
228 * stop state, in which we do not permit the allocation/committing of
229 * any new audit records. We continue to process records but don't
230 * allow any activities that might generate new records. In the
231 * future, we might want to detect when space is available again and
232 * allow operation to continue, but this behavior is sufficient to
233 * meet fail stop requirements in CAPP.
234 */
235 if (audit_fail_stop) {
236 if ((unsigned long)((audit_q_len + audit_pre_q_len + 1) *
237 MAX_AUDIT_RECORD_SIZE) / mnt_stat->f_bsize >=
238 (unsigned long)(mnt_stat->f_bfree)) {
239 if (ppsratecheck(&last_fail, &cur_fail, 1))
240 printf("audit_record_write: free space "
241 "below size of audit queue, failing "
242 "stop\n");
243 audit_in_failure = 1;
244 } else if (audit_in_failure) {
245 /*
246 * Note: if we want to handle recovery, this is the
247 * spot to do it: unset audit_in_failure, and issue a
248 * wakeup on the cv.
249 */
250 }
251 }
252
253 error = vn_rdwr(UIO_WRITE, vp, data, len, (off_t)0, UIO_SYSSPACE,
254 IO_APPEND|IO_UNIT, vfs_context_ucred(ctx), NULL,
255 vfs_context_proc(ctx));
256 if (error == ENOSPC)
257 goto fail_enospc;
258 else if (error)
259 goto fail;
260
261 /*
262 * Catch completion of a queue drain here; if we're draining and the
263 * queue is now empty, fail stop. That audit_fail_stop is implicitly
264 * true, since audit_in_failure can only be set of audit_fail_stop is
265 * set.
266 *
267 * Note: if we handle recovery from audit_in_failure, then we need to
268 * make panic here conditional.
269 */
270 if (audit_in_failure) {
271 if (audit_q_len == 0 && audit_pre_q_len == 0) {
272 (void)VNOP_FSYNC(vp, MNT_WAIT, ctx);
273 panic("Audit store overflow; record queue drained.");
274 }
275 }
276
277 vnode_put(vp);
278 return;
279
280fail_enospc:
281 /*
282 * ENOSPC is considered a special case with respect to failures, as
283 * this can reflect either our preemptive detection of insufficient
284 * space, or ENOSPC returned by the vnode write call.
285 */
286 if (audit_fail_stop) {
287 (void)VNOP_FSYNC(vp, MNT_WAIT, ctx);
288 panic("Audit log space exhausted and fail-stop set.");
289 }
290 (void)audit_send_trigger(AUDIT_TRIGGER_NO_SPACE);
291 audit_suspended = 1;
292
293 /* FALLTHROUGH */
294fail:
295 /*
296 * We have failed to write to the file, so the current record is
297 * lost, which may require an immediate system halt.
298 */
299 if (audit_panic_on_write_fail) {
300 (void)VNOP_FSYNC(vp, MNT_WAIT, ctx);
301 panic("audit_worker: write error %d\n", error);
302 } else if (ppsratecheck(&last_fail, &cur_fail, 1))
303 printf("audit_worker: write error %d\n", error);
304 vnode_put(vp);
305}
306
307/*
308 * Given a kernel audit record, process as required. Kernel audit records
309 * are converted to one, or possibly two, BSM records, depending on whether
310 * there is a user audit record present also. Kernel records need be
311 * converted to BSM before they can be written out. Both types will be
312 * written to disk, and audit pipes.
313 */
314static void
315audit_worker_process_record(struct kaudit_record *ar)
316{
317 struct au_record *bsm;
318 au_class_t class;
319 au_event_t event;
320 au_id_t auid;
321 int error, sorf;
322 int trail_locked;
323
324 /*
325 * We hold the audit_worker_sl lock over both writes, if there are
326 * two, so that the two records won't be split across a rotation and
327 * end up in two different trail files.
328 */
329 if (((ar->k_ar_commit & AR_COMMIT_USER) &&
330 (ar->k_ar_commit & AR_PRESELECT_USER_TRAIL)) ||
331 (ar->k_ar_commit & AR_PRESELECT_TRAIL)) {
332 AUDIT_WORKER_SX_XLOCK();
333 trail_locked = 1;
334 } else
335 trail_locked = 0;
336
337 /*
338 * First, handle the user record, if any: commit to the system trail
339 * and audit pipes as selected.
340 */
341 if ((ar->k_ar_commit & AR_COMMIT_USER) &&
342 (ar->k_ar_commit & AR_PRESELECT_USER_TRAIL)) {
343 AUDIT_WORKER_SX_ASSERT();
344 audit_record_write(audit_vp, &audit_ctx, ar->k_udata,
345 ar->k_ulen);
346 }
347
348 if ((ar->k_ar_commit & AR_COMMIT_USER) &&
349 (ar->k_ar_commit & AR_PRESELECT_USER_PIPE))
350 audit_pipe_submit_user(ar->k_udata, ar->k_ulen);
351
352 if (!(ar->k_ar_commit & AR_COMMIT_KERNEL) ||
353 ((ar->k_ar_commit & AR_PRESELECT_PIPE) == 0 &&
354 (ar->k_ar_commit & AR_PRESELECT_TRAIL) == 0 &&
355 (ar->k_ar_commit & AR_PRESELECT_FILTER) == 0))
356 goto out;
357
358 auid = ar->k_ar.ar_subj_auid;
359 event = ar->k_ar.ar_event;
360 class = au_event_class(event);
361 if (ar->k_ar.ar_errno == 0)
362 sorf = AU_PRS_SUCCESS;
363 else
364 sorf = AU_PRS_FAILURE;
365
366 error = kaudit_to_bsm(ar, &bsm);
367 switch (error) {
368 case BSM_NOAUDIT:
369 goto out;
370
371 case BSM_FAILURE:
372 printf("audit_worker_process_record: BSM_FAILURE\n");
373 goto out;
374
375 case BSM_SUCCESS:
376 break;
377
378 default:
379 panic("kaudit_to_bsm returned %d", error);
380 }
381
382 if (ar->k_ar_commit & AR_PRESELECT_TRAIL) {
383 AUDIT_WORKER_SX_ASSERT();
384 audit_record_write(audit_vp, &audit_ctx, bsm->data, bsm->len);
385 }
386
387 if (ar->k_ar_commit & AR_PRESELECT_PIPE)
388 audit_pipe_submit(auid, event, class, sorf,
389 ar->k_ar_commit & AR_PRESELECT_TRAIL, bsm->data,
390 bsm->len);
391
392 if (ar->k_ar_commit & AR_PRESELECT_FILTER) {
393
394 /*
395 * XXXss - This needs to be generalized so new filters can
396 * be easily plugged in.
397 */
398 audit_sdev_submit(auid, ar->k_ar.ar_subj_asid, bsm->data,
399 bsm->len);
400 }
401
402 kau_free(bsm);
403out:
404 if (trail_locked)
405 AUDIT_WORKER_SX_XUNLOCK();
406}
407
408/*
409 * The audit_worker thread is responsible for watching the event queue,
410 * dequeueing records, converting them to BSM format, and committing them to
411 * disk. In order to minimize lock thrashing, records are dequeued in sets
412 * to a thread-local work queue.
413 *
414 * Note: this means that the effect bound on the size of the pending record
415 * queue is 2x the length of the global queue.
416 */
417__attribute__((noreturn))
418static void
419audit_worker(void)
420{
421 struct kaudit_queue ar_worklist;
422 struct kaudit_record *ar;
423 int lowater_signal;
424
425 if (audit_ctx.vc_thread == NULL)
426 audit_ctx.vc_thread = current_thread();
427
428 TAILQ_INIT(&ar_worklist);
429 mtx_lock(&audit_mtx);
430 while (1) {
431 mtx_assert(&audit_mtx, MA_OWNED);
432
433 /*
434 * Wait for a record.
435 */
436 while (TAILQ_EMPTY(&audit_q))
437 cv_wait_continuation(&audit_worker_cv, &audit_mtx,
438 (thread_continue_t)audit_worker);
439
440 /*
441 * If there are records in the global audit record queue,
442 * transfer them to a thread-local queue and process them
443 * one by one. If we cross the low watermark threshold,
444 * signal any waiting processes that they may wake up and
445 * continue generating records.
446 */
447 lowater_signal = 0;
448 while ((ar = TAILQ_FIRST(&audit_q))) {
449 TAILQ_REMOVE(&audit_q, ar, k_q);
450 audit_q_len--;
451 if (audit_q_len == audit_qctrl.aq_lowater)
452 lowater_signal++;
453 TAILQ_INSERT_TAIL(&ar_worklist, ar, k_q);
454 }
455 if (lowater_signal)
456 cv_broadcast(&audit_watermark_cv);
457
458 mtx_unlock(&audit_mtx);
459 while ((ar = TAILQ_FIRST(&ar_worklist))) {
460 TAILQ_REMOVE(&ar_worklist, ar, k_q);
461 if (ar->k_ar_commit & AR_DRAIN_QUEUE) {
462 audit_q_draining = 0;
463 cv_broadcast(&audit_drain_cv);
464 } else {
465 audit_worker_process_record(ar);
466 audit_free(ar);
467 }
468 }
469 mtx_lock(&audit_mtx);
470 }
471}
472
473/*
474 * audit_rotate_vnode() is called by a user or kernel thread to configure or
475 * de-configure auditing on a vnode. The arguments are the replacement
476 * credential (referenced) and vnode (referenced and opened) to substitute
477 * for the current credential and vnode, if any. If either is set to NULL,
478 * both should be NULL, and this is used to indicate that audit is being
479 * disabled. Any previous cred/vnode will be closed and freed. We re-enable
480 * generating rotation requests to auditd.
481 */
482void
483audit_rotate_vnode(kauth_cred_t cred, struct vnode *vp)
484{
485 kauth_cred_t old_audit_cred;
486 struct vnode *old_audit_vp;
487
488 KASSERT((cred != NULL && vp != NULL) || (cred == NULL && vp == NULL),
489 ("audit_rotate_vnode: cred %p vp %p", cred, vp));
490
491
492 mtx_lock(&audit_mtx);
493 if (audit_enabled && (NULL == vp)) {
494 /* Auditing is currently enabled but will be disabled. */
495
496 /*
497 * Disable auditing now so nothing more is added while the
498 * audit worker thread is draining the audit record queue.
499 */
500 audit_enabled = 0;
501
502 /*
503 * Drain the auditing queue by inserting a drain record at the
504 * end of the queue and waiting for the audit worker thread
505 * to find this record and signal that it is done before
506 * we close the audit trail.
507 */
508 audit_q_draining = 1;
509 while (audit_q_len >= audit_qctrl.aq_hiwater)
510 cv_wait(&audit_watermark_cv, &audit_mtx);
511 TAILQ_INSERT_TAIL(&audit_q, &audit_drain_kar, k_q);
512 audit_q_len++;
513 cv_signal(&audit_worker_cv);
514 }
515
516 /* If the audit queue is draining then wait here until it's done. */
517 while (audit_q_draining)
518 cv_wait(&audit_drain_cv, &audit_mtx);
519 mtx_unlock(&audit_mtx);
520
521
522 /*
523 * Rotate the vnode/cred, and clear the rotate flag so that we will
524 * send a rotate trigger if the new file fills.
525 */
526 AUDIT_WORKER_SX_XLOCK();
527 old_audit_cred = audit_ctx.vc_ucred;
528 old_audit_vp = audit_vp;
529 audit_ctx.vc_ucred = cred;
530 audit_vp = vp;
531 audit_file_rotate_wait = 0;
532 audit_enabled = (audit_vp != NULL);
533 AUDIT_WORKER_SX_XUNLOCK();
534
535 /*
536 * If there was an old vnode/credential, close and free.
537 */
538 if (old_audit_vp != NULL) {
539 if (vnode_get(old_audit_vp) == 0) {
540 vn_close(old_audit_vp, AUDIT_CLOSE_FLAGS,
541 vfs_context_kernel());
542 vnode_put(old_audit_vp);
543 } else
544 printf("audit_rotate_vnode: Couldn't close "
545 "audit file.\n");
546 kauth_cred_unref(&old_audit_cred);
547 }
548}
549
550void
551audit_worker_init(void)
552{
553
554 AUDIT_WORKER_SX_INIT();
555 kernel_thread_start((thread_continue_t)audit_worker, NULL,
556 &audit_thread);
557 if (audit_thread == THREAD_NULL)
558 panic("audit_worker_init: Couldn't create audit_worker thread");
559}
560