61#include <bsm/audit_kevents.h>
62
63#include <netinet/in.h>
64#include <netinet/in_pcb.h>
65
66#include <security/audit/audit.h>
67#include <security/audit/audit_private.h>
68
69#include <vm/uma.h>
70
71/*
72 * The AUDIT_EXCESSIVELY_VERBOSE define enables a number of
73 * gratuitously noisy printf's to the console. Due to the
74 * volume, it should be left off unless you want your system
75 * to churn a lot whenever the audit record flow gets high.
76 */
77//#define AUDIT_EXCESSIVELY_VERBOSE
78#ifdef AUDIT_EXCESSIVELY_VERBOSE
79#define AUDIT_PRINTF(x) printf x
80#else
81#define AUDIT_PRINTF(X)
82#endif
83
84static uma_zone_t audit_record_zone;
85static MALLOC_DEFINE(M_AUDITPROC, "audit_proc", "Audit process storage");
86MALLOC_DEFINE(M_AUDITDATA, "audit_data", "Audit data storage");
87MALLOC_DEFINE(M_AUDITPATH, "audit_path", "Audit path storage");
88MALLOC_DEFINE(M_AUDITTEXT, "audit_text", "Audit text storage");
89
90/*
91 * Audit control settings that are set/read by system calls and are
92 * hence non-static.
93 */
94/*
95 * Define the audit control flags.
96 */
97int audit_enabled;
98int audit_suspended;
99
100/*
101 * Flags controlling behavior in low storage situations.
102 * Should we panic if a write fails? Should we fail stop
103 * if we're out of disk space?
104 */
105int audit_panic_on_write_fail;
106int audit_fail_stop;
107
108/*
109 * Are we currently "failing stop" due to out of disk space?
110 */
111static int audit_in_failure;
112
113/*
114 * Global audit statistiscs.
115 */
116struct audit_fstat audit_fstat;
117
118/*
119 * Preselection mask for non-attributable events.
120 */
121struct au_mask audit_nae_mask;
122
123/*
124 * Mutex to protect global variables shared between various threads and
125 * processes.
126 */
127static struct mtx audit_mtx;
128
129/*
130 * Queue of audit records ready for delivery to disk. We insert new
131 * records at the tail, and remove records from the head. Also,
132 * a count of the number of records used for checking queue depth.
133 * In addition, a counter of records that we have allocated but are
134 * not yet in the queue, which is needed to estimate the total
135 * size of the combined set of records outstanding in the system.
136 */
137static TAILQ_HEAD(, kaudit_record) audit_q;
138static int audit_q_len;
139static int audit_pre_q_len;
140
141/*
142 * Audit queue control settings (minimum free, low/high water marks, etc.)
143 */
144struct au_qctrl audit_qctrl;
145
146/*
147 * Condition variable to signal to the worker that it has work to do:
148 * either new records are in the queue, or a log replacement is taking
149 * place.
150 */
151static struct cv audit_cv;
152
153/*
154 * Worker thread that will schedule disk I/O, etc.
155 */
156static struct proc *audit_thread;
157
158/*
159 * When an audit log is rotated, the actual rotation must be performed
160 * by the audit worker thread, as it may have outstanding writes on the
161 * current audit log. audit_replacement_vp holds the vnode replacing
162 * the current vnode. We can't let more than one replacement occur
163 * at a time, so if more than one thread requests a replacement, only
164 * one can have the replacement "in progress" at any given moment. If
165 * a thread tries to replace the audit vnode and discovers a replacement
166 * is already in progress (i.e., audit_replacement_flag != 0), then it
167 * will sleep on audit_replacement_cv waiting its turn to perform a
168 * replacement. When a replacement is completed, this cv is signalled
169 * by the worker thread so a waiting thread can start another replacement.
170 * We also store a credential to perform audit log write operations with.
171 *
172 * The current credential and vnode are thread-local to audit_worker.
173 */
174static struct cv audit_replacement_cv;
175
176static int audit_replacement_flag;
177static struct vnode *audit_replacement_vp;
178static struct ucred *audit_replacement_cred;
179
180/*
181 * Condition variable to signal to the worker that it has work to do:
182 * either new records are in the queue, or a log replacement is taking
183 * place.
184 */
185static struct cv audit_commit_cv;
186
187/*
188 * Condition variable for auditing threads wait on when in fail-stop mode.
189 * Threads wait on this CV forever (and ever), never seeing the light of
190 * day again.
191 */
192static struct cv audit_fail_cv;
193
194/*
195 * Flags related to Kernel->user-space communication.
196 */
197static int audit_file_rotate_wait;
198
199/*
200 * Construct an audit record for the passed thread.
201 */
202static int
203audit_record_ctor(void *mem, int size, void *arg, int flags)
204{
205 struct kaudit_record *ar;
206 struct thread *td;
207
208 KASSERT(sizeof(*ar) == size, ("audit_record_ctor: wrong size"));
209
210 td = arg;
211 ar = mem;
212 bzero(ar, sizeof(*ar));
213 ar->k_ar.ar_magic = AUDIT_RECORD_MAGIC;
214 nanotime(&ar->k_ar.ar_starttime);
215
216 /*
217 * Export the subject credential.
218 *
219 * XXXAUDIT: td_ucred access is OK without proc lock, but some other
220 * fields here may require the proc lock.
221 */
222 cru2x(td->td_ucred, &ar->k_ar.ar_subj_cred);
223 ar->k_ar.ar_subj_ruid = td->td_ucred->cr_ruid;
224 ar->k_ar.ar_subj_rgid = td->td_ucred->cr_rgid;
225 ar->k_ar.ar_subj_egid = td->td_ucred->cr_groups[0];
226 ar->k_ar.ar_subj_auid = td->td_proc->p_au->ai_auid;
227 ar->k_ar.ar_subj_asid = td->td_proc->p_au->ai_asid;
228 ar->k_ar.ar_subj_pid = td->td_proc->p_pid;
229 ar->k_ar.ar_subj_amask = td->td_proc->p_au->ai_mask;
230 ar->k_ar.ar_subj_term = td->td_proc->p_au->ai_termid;
231 bcopy(td->td_proc->p_comm, ar->k_ar.ar_subj_comm, MAXCOMLEN);
232
233 return (0);
234}
235
236static void
237audit_record_dtor(void *mem, int size, void *arg)
238{
239 struct kaudit_record *ar;
240
241 KASSERT(sizeof(*ar) == size, ("audit_record_dtor: wrong size"));
242
243 ar = mem;
244 if (ar->k_ar.ar_arg_upath1 != NULL)
245 free(ar->k_ar.ar_arg_upath1, M_AUDITPATH);
246 if (ar->k_ar.ar_arg_upath2 != NULL)
247 free(ar->k_ar.ar_arg_upath2, M_AUDITPATH);
248 if (ar->k_ar.ar_arg_text != NULL)
249 free(ar->k_ar.ar_arg_text, M_AUDITTEXT);
250 if (ar->k_udata != NULL)
251 free(ar->k_udata, M_AUDITDATA);
252}
253
254/*
255 * XXXAUDIT: Should adjust comments below to make it clear that we get to
256 * this point only if we believe we have storage, so not having space here
257 * is a violation of invariants derived from administrative procedures.
258 * I.e., someone else has written to the audit partition, leaving less space
259 * than we accounted for.
260 */
261static int
262audit_record_write(struct vnode *vp, struct kaudit_record *ar,
263 struct ucred *cred, struct thread *td)
264{
265 int ret;
266 long temp;
267 struct au_record *bsm;
268 struct vattr vattr;
269 struct statfs *mnt_stat = &vp->v_mount->mnt_stat;
270 int vfslocked;
271
272 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
273
274 /*
275 * First, gather statistics on the audit log file and file system
276 * so that we know how we're doing on space. In both cases,
277 * if we're unable to perform the operation, we drop the record
278 * and return. However, this is arguably an assertion failure.
279 * XXX Need a FreeBSD equivalent.
280 */
281 ret = VFS_STATFS(vp->v_mount, mnt_stat, td);
282 if (ret)
283 goto out;
284
285 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
286 ret = VOP_GETATTR(vp, &vattr, cred, td);
287 VOP_UNLOCK(vp, 0, td);
288 if (ret)
289 goto out;
290
291 /* update the global stats struct */
292 audit_fstat.af_currsz = vattr.va_size;
293
294 /*
295 * XXX Need to decide what to do if the trigger to the audit daemon
296 * fails.
297 */
298
299 /*
300 * If we fall below minimum free blocks (hard limit), tell the audit
301 * daemon to force a rotation off of the file system. We also stop
302 * writing, which means this audit record is probably lost.
303 * If we fall below the minimum percent free blocks (soft limit),
304 * then kindly suggest to the audit daemon to do something.
305 */
306 if (mnt_stat->f_bfree < AUDIT_HARD_LIMIT_FREE_BLOCKS) {
307 send_trigger(AUDIT_TRIGGER_NO_SPACE);
308 /* Hopefully userspace did something about all the previous
309 * triggers that were sent prior to this critical condition.
310 * If fail-stop is set, then we're done; goodnight Gracie.
311 */
312 if (audit_fail_stop)
313 panic("Audit log space exhausted and fail-stop set.");
314 else {
315 audit_suspended = 1;
316 ret = ENOSPC;
317 goto out;
318 }
319 } else
320 /*
321 * Send a message to the audit daemon that disk space
322 * is getting low.
323 *
324 * XXXAUDIT: Check math and block size calculation here.
325 */
326 if (audit_qctrl.aq_minfree != 0) {
327 temp = mnt_stat->f_blocks / (100 /
328 audit_qctrl.aq_minfree);
329 if (mnt_stat->f_bfree < temp)
330 send_trigger(AUDIT_TRIGGER_LOW_SPACE);
331 }
332
333 /* Check if the current log file is full; if so, call for
334 * a log rotate. This is not an exact comparison; we may
335 * write some records over the limit. If that's not
336 * acceptable, then add a fudge factor here.
337 */
338 if ((audit_fstat.af_filesz != 0) &&
339 (audit_file_rotate_wait == 0) &&
340 (vattr.va_size >= audit_fstat.af_filesz)) {
341 audit_file_rotate_wait = 1;
342 send_trigger(AUDIT_TRIGGER_OPEN_NEW);
343 }
344
345 /*
346 * If the estimated amount of audit data in the audit event queue
347 * (plus records allocated but not yet queued) has reached the
348 * amount of free space on the disk, then we need to go into an
349 * audit fail stop state, in which we do not permit the
350 * allocation/committing of any new audit records. We continue to
351 * process packets but don't allow any activities that might
352 * generate new records. In the future, we might want to detect
353 * when space is available again and allow operation to continue,
354 * but this behavior is sufficient to meet fail stop requirements
355 * in CAPP.
356 */
357 if (audit_fail_stop &&
358 (unsigned long)
359 ((audit_q_len + audit_pre_q_len + 1) * MAX_AUDIT_RECORD_SIZE) /
360 mnt_stat->f_bsize >= (unsigned long)(mnt_stat->f_bfree)) {
361 printf(
362 "audit_worker: free space below size of audit queue, failing stop\n");
363 audit_in_failure = 1;
364 }
365
366 /*
367 * If there is a user audit record attached to the kernel record,
368 * then write the user record.
369 */
370 /* XXX Need to decide a few things here: IF the user audit
371 * record is written, but the write of the kernel record fails,
372 * what to do? Should the kernel record come before or after the
373 * user record? For now, we write the user record first, and
374 * we ignore errors.
375 */
376 if (ar->k_ar_commit & AR_COMMIT_USER) {
377 /*
378 * Try submitting the record to any active audit pipes.
379 */
380 audit_pipe_submit((void *)ar->k_udata, ar->k_ulen);
381
382 /*
383 * And to disk.
384 */
385 ret = vn_rdwr(UIO_WRITE, vp, (void *)ar->k_udata, ar->k_ulen,
386 (off_t)0, UIO_SYSSPACE, IO_APPEND|IO_UNIT, cred, NULL,
387 NULL, td);
388 if (ret)
389 goto out;
390 }
391
392 /*
393 * Convert the internal kernel record to BSM format and write it
394 * out if everything's OK.
395 */
396 if (!(ar->k_ar_commit & AR_COMMIT_KERNEL)) {
397 ret = 0;
398 goto out;
399 }
400
401 /*
402 * XXXAUDIT: Should we actually allow this conversion to fail? With
403 * sleeping memory allocation and invariants checks, perhaps not.
404 */
405 ret = kaudit_to_bsm(ar, &bsm);
406 if (ret == BSM_NOAUDIT) {
407 ret = 0;
408 goto out;
409 }
410
411 /*
412 * XXX: We drop the record on BSM conversion failure, but really
413 * this is an assertion failure.
414 */
415 if (ret == BSM_FAILURE) {
416 AUDIT_PRINTF(("BSM conversion failure\n"));
417 ret = EINVAL;
418 goto out;
419 }
420
421 /*
422 * Try submitting the record to any active audit pipes.
423 */
424 audit_pipe_submit((void *)bsm->data, bsm->len);
425
426 /*
427 * XXX
428 * We should break the write functionality away from the BSM record
429 * generation and have the BSM generation done before this function
430 * is called. This function will then take the BSM record as a
431 * parameter.
432 */
433 ret = (vn_rdwr(UIO_WRITE, vp, (void *)bsm->data, bsm->len,
434 (off_t)0, UIO_SYSSPACE, IO_APPEND|IO_UNIT, cred, NULL, NULL, td));
435
436 kau_free(bsm);
437
438out:
439 /*
440 * When we're done processing the current record, we have to
441 * check to see if we're in a failure mode, and if so, whether
442 * this was the last record left to be drained. If we're done
443 * draining, then we fsync the vnode and panic.
444 */
445 if (audit_in_failure &&
446 audit_q_len == 0 && audit_pre_q_len == 0) {
447 VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td);
448 (void)VOP_FSYNC(vp, MNT_WAIT, td);
449 VOP_UNLOCK(vp, 0, td);
450 panic("Audit store overflow; record queue drained.");
451 }
452
453 VFS_UNLOCK_GIANT(vfslocked);
454
455 return (ret);
456}
457
458/*
459 * The audit_worker thread is responsible for watching the event queue,
460 * dequeueing records, converting them to BSM format, and committing them to
461 * disk. In order to minimize lock thrashing, records are dequeued in sets
462 * to a thread-local work queue. In addition, the audit_work performs the
463 * actual exchange of audit log vnode pointer, as audit_vp is a thread-local
464 * variable.
465 */
466static void
467audit_worker(void *arg)
468{
469 int do_replacement_signal, error;
470 TAILQ_HEAD(, kaudit_record) ar_worklist;
471 struct kaudit_record *ar;
472 struct vnode *audit_vp, *old_vp;
473 int vfslocked;
474
475 struct ucred *audit_cred, *old_cred;
476 struct thread *audit_td;
477
478 AUDIT_PRINTF(("audit_worker starting\n"));
479
480 /*
481 * These are thread-local variables requiring no synchronization.
482 */
483 TAILQ_INIT(&ar_worklist);
484 audit_cred = NULL;
485 audit_td = curthread;
486 audit_vp = NULL;
487
488 mtx_lock(&audit_mtx);
489 while (1) {
490 /*
491 * First priority: replace the audit log target if requested.
492 * Accessing the vnode here requires dropping the audit_mtx;
493 * in case another replacement was scheduled while the mutex
494 * was released, we loop.
495 *
496 * XXX It could well be we should drain existing records
497 * first to ensure that the timestamps and ordering
498 * are right.
499 */
500 do_replacement_signal = 0;
501 while (audit_replacement_flag != 0) {
502 old_cred = audit_cred;
503 old_vp = audit_vp;
504 audit_cred = audit_replacement_cred;
505 audit_vp = audit_replacement_vp;
506 audit_replacement_cred = NULL;
507 audit_replacement_vp = NULL;
508 audit_replacement_flag = 0;
509
510 audit_enabled = (audit_vp != NULL);
511
512 /*
513 * XXX: What to do about write failures here?
514 */
515 if (old_vp != NULL) {
516 AUDIT_PRINTF(("Closing old audit file\n"));
517 mtx_unlock(&audit_mtx);
518 vfslocked = VFS_LOCK_GIANT(old_vp->v_mount);
519 vn_close(old_vp, AUDIT_CLOSE_FLAGS, old_cred,
520 audit_td);
521 VFS_UNLOCK_GIANT(vfslocked);
522 crfree(old_cred);
523 mtx_lock(&audit_mtx);
524 old_cred = NULL;
525 old_vp = NULL;
526 AUDIT_PRINTF(("Audit file closed\n"));
527 }
528 if (audit_vp != NULL) {
529 AUDIT_PRINTF(("Opening new audit file\n"));
530 }
531 do_replacement_signal = 1;
532 }
533 /*
534 * Signal that replacement have occurred to wake up and
535 * start any other replacements started in parallel. We can
536 * continue about our business in the mean time. We
537 * broadcast so that both new replacements can be inserted,
538 * but also so that the source(s) of replacement can return
539 * successfully.
540 */
541 if (do_replacement_signal)
542 cv_broadcast(&audit_replacement_cv);
543
544 /*
545 * Next, check to see if we have any records to drain into
546 * the vnode. If not, go back to waiting for an event.
547 */
548 if (TAILQ_EMPTY(&audit_q)) {
549 AUDIT_PRINTF(("audit_worker waiting\n"));
550 cv_wait(&audit_cv, &audit_mtx);
551 AUDIT_PRINTF(("audit_worker woken up\n"));
552 AUDIT_PRINTF(("audit_worker: new vp = %p; value of flag %d\n",
553 audit_replacement_vp, audit_replacement_flag));
554 continue;
555 }
556
557 /*
558 * If we have records, but there's no active vnode to write
559 * to, drain the record queue. Generally, we prevent the
560 * unnecessary allocation of records elsewhere, but we need
561 * to allow for races between conditional allocation and
562 * queueing. Go back to waiting when we're done.
563 */
564 if (audit_vp == NULL) {
565 while ((ar = TAILQ_FIRST(&audit_q))) {
566 TAILQ_REMOVE(&audit_q, ar, k_q);
567 uma_zfree(audit_record_zone, ar);
568 audit_q_len--;
569 /*
570 * XXXRW: Why broadcast if we hold the
571 * mutex and know that audit_vp is NULL?
572 */
573 if (audit_q_len <= audit_qctrl.aq_lowater)
574 cv_broadcast(&audit_commit_cv);
575 }
576 continue;
577 }
578
579 /*
580 * We have both records to write and an active vnode to write
581 * to. Dequeue a record, and start the write. Eventually,
582 * it might make sense to dequeue several records and perform
583 * our own clustering, if the lower layers aren't doing it
584 * automatically enough.
585 */
586 while ((ar = TAILQ_FIRST(&audit_q))) {
587 TAILQ_REMOVE(&audit_q, ar, k_q);
588 audit_q_len--;
589 if (audit_q_len <= audit_qctrl.aq_lowater)
590 cv_broadcast(&audit_commit_cv);
591 TAILQ_INSERT_TAIL(&ar_worklist, ar, k_q);
592 }
593
594 mtx_unlock(&audit_mtx);
595 while ((ar = TAILQ_FIRST(&ar_worklist))) {
596 TAILQ_REMOVE(&ar_worklist, ar, k_q);
597 if (audit_vp != NULL) {
598 error = audit_record_write(audit_vp, ar,
599 audit_cred, audit_td);
600 if (error && audit_panic_on_write_fail)
601 panic("audit_worker: write error %d\n",
602 error);
603 else if (error)
604 printf("audit_worker: write error %d\n",
605 error);
606 }
607 uma_zfree(audit_record_zone, ar);
608 }
609 mtx_lock(&audit_mtx);
610 }
611}
612
613/*
614 * Initialize the Audit subsystem: configuration state, work queue,
615 * synchronization primitives, worker thread, and trigger device node. Also
616 * call into the BSM assembly code to initialize it.
617 */
618static void
619audit_init(void)
620{
621 int error;
622
623 printf("Security auditing service present\n");
624 audit_enabled = 0;
625 audit_suspended = 0;
626 audit_panic_on_write_fail = 0;
627 audit_fail_stop = 0;
628 audit_in_failure = 0;
629
630 audit_replacement_vp = NULL;
631 audit_replacement_cred = NULL;
632 audit_replacement_flag = 0;
633
634 audit_fstat.af_filesz = 0; /* '0' means unset, unbounded */
635 audit_fstat.af_currsz = 0;
636 audit_nae_mask.am_success = AU_NULL;
637 audit_nae_mask.am_failure = AU_NULL;
638
639 TAILQ_INIT(&audit_q);
640 audit_q_len = 0;
641 audit_pre_q_len = 0;
642 audit_qctrl.aq_hiwater = AQ_HIWATER;
643 audit_qctrl.aq_lowater = AQ_LOWATER;
644 audit_qctrl.aq_bufsz = AQ_BUFSZ;
645 audit_qctrl.aq_minfree = AU_FS_MINFREE;
646
647 mtx_init(&audit_mtx, "audit_mtx", NULL, MTX_DEF);
648 cv_init(&audit_cv, "audit_cv");
649 cv_init(&audit_replacement_cv, "audit_replacement_cv");
650 cv_init(&audit_commit_cv, "audit_commit_cv");
651 cv_init(&audit_fail_cv, "audit_fail_cv");
652
653 audit_record_zone = uma_zcreate("audit_record_zone",
654 sizeof(struct kaudit_record), audit_record_ctor,
655 audit_record_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
656
657 /* Initialize the BSM audit subsystem. */
658 kau_init();
659
660 audit_file_rotate_wait = 0;
661 audit_trigger_init();
662
663 /* Register shutdown handler. */
664 EVENTHANDLER_REGISTER(shutdown_pre_sync, audit_shutdown, NULL,
665 SHUTDOWN_PRI_FIRST);
666
667 error = kthread_create(audit_worker, NULL, &audit_thread, RFHIGHPID,
668 0, "audit_worker");
669 if (error != 0)
670 panic("audit_init: kthread_create returned %d", error);
671}
672
673SYSINIT(audit_init, SI_SUB_AUDIT, SI_ORDER_FIRST, audit_init, NULL)
674
675/*
676 * audit_rotate_vnode() is called by a user or kernel thread to configure or
677 * de-configure auditing on a vnode. The arguments are the replacement
678 * credential and vnode to substitute for the current credential and vnode,
679 * if any. If either is set to NULL, both should be NULL, and this is used
680 * to indicate that audit is being disabled. The real work is done in the
681 * audit_worker thread, but audit_rotate_vnode() waits synchronously for that
682 * to complete.
683 *
684 * The vnode should be referenced and opened by the caller. The credential
685 * should be referenced. audit_rotate_vnode() will own both references as of
686 * this call, so the caller should not release either.
687 *
688 * XXXAUDIT: Review synchronize communication logic. Really, this is a
689 * message queue of depth 1.
690 *
691 * XXXAUDIT: Enhance the comments below to indicate that we are basically
692 * acquiring ownership of the communications queue, inserting our message,
693 * and waiting for an acknowledgement.
694 */
695void
696audit_rotate_vnode(struct ucred *cred, struct vnode *vp)
697{
698
699 /*
700 * If other parallel log replacements have been requested, we wait
701 * until they've finished before continuing.
702 */
703 mtx_lock(&audit_mtx);
704 while (audit_replacement_flag != 0) {
705 AUDIT_PRINTF(("audit_rotate_vnode: sleeping to wait for "
706 "flag\n"));
707 cv_wait(&audit_replacement_cv, &audit_mtx);
708 AUDIT_PRINTF(("audit_rotate_vnode: woken up (flag %d)\n",
709 audit_replacement_flag));
710 }
711 audit_replacement_cred = cred;
712 audit_replacement_flag = 1;
713 audit_replacement_vp = vp;
714
715 /*
716 * Wake up the audit worker to perform the exchange once we
717 * release the mutex.
718 */
719 cv_signal(&audit_cv);
720
721 /*
722 * Wait for the audit_worker to broadcast that a replacement has
723 * taken place; we know that once this has happened, our vnode
724 * has been replaced in, so we can return successfully.
725 */
726 AUDIT_PRINTF(("audit_rotate_vnode: waiting for news of "
727 "replacement\n"));
728 cv_wait(&audit_replacement_cv, &audit_mtx);
729 AUDIT_PRINTF(("audit_rotate_vnode: change acknowledged by "
730 "audit_worker (flag " "now %d)\n", audit_replacement_flag));
731 mtx_unlock(&audit_mtx);
732
733 audit_file_rotate_wait = 0; /* We can now request another rotation */
734}
735
736/*
737 * Drain the audit queue and close the log at shutdown. Note that this can
738 * be called both from the system shutdown path and also from audit
739 * configuration syscalls, so 'arg' and 'howto' are ignored.
740 */
741void
742audit_shutdown(void *arg, int howto)
743{
744
745 audit_rotate_vnode(NULL, NULL);
746}
747
748/*
749 * Return the current thread's audit record, if any.
750 */
751__inline__ struct kaudit_record *
752currecord(void)
753{
754
755 return (curthread->td_ar);
756}
757
758/*
759 * MPSAFE
760 *
761 * XXXAUDIT: There are a number of races present in the code below due to
762 * release and re-grab of the mutex. The code should be revised to become
763 * slightly less racy.
764 *
765 * XXXAUDIT: Shouldn't there be logic here to sleep waiting on available
766 * pre_q space, suspending the system call until there is room?
767 */
768struct kaudit_record *
769audit_new(int event, struct thread *td)
770{
771 struct kaudit_record *ar;
772 int no_record;
773
774 mtx_lock(&audit_mtx);
775 no_record = (audit_suspended || !audit_enabled);
776 mtx_unlock(&audit_mtx);
777 if (no_record)
778 return (NULL);
779
780 /*
781 * XXX: The number of outstanding uncommitted audit records is
782 * limited to the number of concurrent threads servicing system
783 * calls in the kernel.
784 */
785 ar = uma_zalloc_arg(audit_record_zone, td, M_WAITOK);
786 ar->k_ar.ar_event = event;
787
788 mtx_lock(&audit_mtx);
789 audit_pre_q_len++;
790 mtx_unlock(&audit_mtx);
791
792 return (ar);
793}
794
795/*
796 * MPSAFE
797 */
798void
799audit_commit(struct kaudit_record *ar, int error, int retval)
800{
801 int sorf;
802 struct au_mask *aumask;
803
804 if (ar == NULL)
805 return;
806
807 /*
808 * Decide whether to commit the audit record by checking the
809 * error value from the system call and using the appropriate
810 * audit mask.
811 *
812 * XXXAUDIT: Synchronize access to audit_nae_mask?
813 */
814 if (ar->k_ar.ar_subj_auid == AU_DEFAUDITID)
815 aumask = &audit_nae_mask;
816 else
817 aumask = &ar->k_ar.ar_subj_amask;
818
819 if (error)
820 sorf = AU_PRS_FAILURE;
821 else
822 sorf = AU_PRS_SUCCESS;
823
824 switch(ar->k_ar.ar_event) {
825
826 case AUE_OPEN_RWTC:
827 /* The open syscall always writes a AUE_OPEN_RWTC event; change
828 * it to the proper type of event based on the flags and the
829 * error value.
830 */
831 ar->k_ar.ar_event = flags_and_error_to_openevent(
832 ar->k_ar.ar_arg_fflags, error);
833 break;
834
835 case AUE_SYSCTL:
836 ar->k_ar.ar_event = ctlname_to_sysctlevent(
837 ar->k_ar.ar_arg_ctlname, ar->k_ar.ar_valid_arg);
838 break;
839
840 case AUE_AUDITON:
841 /* Convert the auditon() command to an event */
842 ar->k_ar.ar_event = auditon_command_event(ar->k_ar.ar_arg_cmd);
843 break;
844 }
845
846 if (au_preselect(ar->k_ar.ar_event, aumask, sorf) != 0)
847 ar->k_ar_commit |= AR_COMMIT_KERNEL;
848
849 /*
850 * XXXRW: Why is this necessary? Should we ever accept a record that
851 * we're not willing to commit?
852 */
853 if ((ar->k_ar_commit & (AR_COMMIT_USER | AR_COMMIT_KERNEL)) == 0) {
854 mtx_lock(&audit_mtx);
855 audit_pre_q_len--;
856 mtx_unlock(&audit_mtx);
857 uma_zfree(audit_record_zone, ar);
858 return;
859 }
860
861 ar->k_ar.ar_errno = error;
862 ar->k_ar.ar_retval = retval;
863
864 /*
865 * We might want to do some system-wide post-filtering
866 * here at some point.
867 */
868
869 /*
870 * Timestamp system call end.
871 */
872 nanotime(&ar->k_ar.ar_endtime);
873
874 mtx_lock(&audit_mtx);
875
876 /*
877 * Note: it could be that some records initiated while audit was
878 * enabled should still be committed?
879 */
880 if (audit_suspended || !audit_enabled) {
881 audit_pre_q_len--;
882 mtx_unlock(&audit_mtx);
883 uma_zfree(audit_record_zone, ar);
884 return;
885 }
886
887 /*
888 * Constrain the number of committed audit records based on
889 * the configurable parameter.
890 */
891 while (audit_q_len >= audit_qctrl.aq_hiwater) {
892 AUDIT_PRINTF(("audit_commit: sleeping to wait for "
893 "audit queue to drain below high water mark\n"));
894 cv_wait(&audit_commit_cv, &audit_mtx);
895 AUDIT_PRINTF(("audit_commit: woke up waiting for "
896 "audit queue draining\n"));
897 }
898
899 TAILQ_INSERT_TAIL(&audit_q, ar, k_q);
900 audit_q_len++;
901 audit_pre_q_len--;
902 cv_signal(&audit_cv);
903 mtx_unlock(&audit_mtx);
904}
905
906/*
907 * audit_syscall_enter() is called on entry to each system call. It is
908 * responsible for deciding whether or not to audit the call (preselection),
909 * and if so, allocating a per-thread audit record. audit_new() will fill in
910 * basic thread/credential properties.
911 */
912void
913audit_syscall_enter(unsigned short code, struct thread *td)
914{
915 int audit_event;
916 struct au_mask *aumask;
917
918 KASSERT(td->td_ar == NULL, ("audit_syscall_enter: td->td_ar != NULL"));
919
920 /*
921 * In FreeBSD, each ABI has its own system call table, and hence
922 * mapping of system call codes to audit events. Convert the code to
923 * an audit event identifier using the process system call table
924 * reference. In Darwin, there's only one, so we use the global
925 * symbol for the system call table.
926 *
927 * XXXAUDIT: Should we audit that a bad system call was made, and if
928 * so, how?
929 */
930 if (code >= td->td_proc->p_sysent->sv_size)
931 return;
932
933 audit_event = td->td_proc->p_sysent->sv_table[code].sy_auevent;
934 if (audit_event == AUE_NULL)
935 return;
936
937 /*
938 * Check which audit mask to use; either the kernel non-attributable
939 * event mask or the process audit mask.
940 */
941 if (td->td_proc->p_au->ai_auid == AU_DEFAUDITID)
942 aumask = &audit_nae_mask;
943 else
944 aumask = &td->td_proc->p_au->ai_mask;
945
946 /*
947 * Allocate an audit record, if preselection allows it, and store
948 * in the thread for later use.
949 */
950 if (au_preselect(audit_event, aumask,
951 AU_PRS_FAILURE | AU_PRS_SUCCESS)) {
952 /*
953 * If we're out of space and need to suspend unprivileged
954 * processes, do that here rather than trying to allocate
955 * another audit record.
956 *
957 * XXXRW: We might wish to be able to continue here in the
958 * future, if the system recovers. That should be possible
959 * by means of checking the condition in a loop around
960 * cv_wait(). It might be desirable to reevaluate whether an
961 * audit record is still required for this event by
962 * re-calling au_preselect().
963 */
964 if (audit_in_failure && suser(td) != 0) {
965 cv_wait(&audit_fail_cv, &audit_mtx);
966 panic("audit_failing_stop: thread continued");
967 }
968 td->td_ar = audit_new(audit_event, td);
969 } else
970 td->td_ar = NULL;
971}
972
973/*
974 * audit_syscall_exit() is called from the return of every system call, or in
975 * the event of exit1(), during the execution of exit1(). It is responsible
976 * for committing the audit record, if any, along with return condition.
977 */
978void
979audit_syscall_exit(int error, struct thread *td)
980{
981 int retval;
982
983 /*
984 * Commit the audit record as desired; once we pass the record
985 * into audit_commit(), the memory is owned by the audit
986 * subsystem.
987 * The return value from the system call is stored on the user
988 * thread. If there was an error, the return value is set to -1,
989 * imitating the behavior of the cerror routine.
990 */
991 if (error)
992 retval = -1;
993 else
994 retval = td->td_retval[0];
995
996 audit_commit(td->td_ar, error, retval);
997 if (td->td_ar != NULL)
998 AUDIT_PRINTF(("audit record committed by pid %d\n",
999 td->td_proc->p_pid));
1000 td->td_ar = NULL;
1001
1002}
1003
1004/*
1005 * Allocate storage for a new process (init, or otherwise).
1006 */
1007void
1008audit_proc_alloc(struct proc *p)
1009{
1010
1011 KASSERT(p->p_au == NULL, ("audit_proc_alloc: p->p_au != NULL (%d)",
1012 p->p_pid));
1013 p->p_au = malloc(sizeof(*(p->p_au)), M_AUDITPROC, M_WAITOK);
1014 /* XXXAUDIT: Zero? Slab allocate? */
1015 //printf("audit_proc_alloc: pid %d p_au %p\n", p->p_pid, p->p_au);
1016}
1017
1018/*
1019 * Allocate storage for a new thread.
1020 */
1021void
1022audit_thread_alloc(struct thread *td)
1023{
1024
1025 td->td_ar = NULL;
1026}
1027
1028/*
1029 * Thread destruction.
1030 */
1031void
1032audit_thread_free(struct thread *td)
1033{
1034
1035 KASSERT(td->td_ar == NULL, ("audit_thread_free: td_ar != NULL"));
1036}
1037
1038/*
1039 * Initialize the audit information for the a process, presumably the first
1040 * process in the system.
1041 * XXX It is not clear what the initial values should be for audit ID,
1042 * session ID, etc.
1043 */
1044void
1045audit_proc_kproc0(struct proc *p)
1046{
1047
1048 KASSERT(p->p_au != NULL, ("audit_proc_kproc0: p->p_au == NULL (%d)",
1049 p->p_pid));
1050 //printf("audit_proc_kproc0: pid %d p_au %p\n", p->p_pid, p->p_au);
1051 bzero(p->p_au, sizeof(*(p)->p_au));
1052}
1053
1054void
1055audit_proc_init(struct proc *p)
1056{
1057
1058 KASSERT(p->p_au != NULL, ("audit_proc_init: p->p_au == NULL (%d)",
1059 p->p_pid));
1060 //printf("audit_proc_init: pid %d p_au %p\n", p->p_pid, p->p_au);
1061 bzero(p->p_au, sizeof(*(p)->p_au));
1062 p->p_au->ai_auid = AU_DEFAUDITID;
1063}
1064
1065/*
1066 * Copy the audit info from the parent process to the child process when
1067 * a fork takes place.
1068 */
1069void
1070audit_proc_fork(struct proc *parent, struct proc *child)
1071{
1072
1073 PROC_LOCK_ASSERT(parent, MA_OWNED);
1074 PROC_LOCK_ASSERT(child, MA_OWNED);
1075 KASSERT(parent->p_au != NULL,
1076 ("audit_proc_fork: parent->p_au == NULL (%d)", parent->p_pid));
1077 KASSERT(child->p_au != NULL,
1078 ("audit_proc_fork: child->p_au == NULL (%d)", child->p_pid));
1079 //printf("audit_proc_fork: parent pid %d p_au %p\n", parent->p_pid,
1080 // parent->p_au);
1081 //printf("audit_proc_fork: child pid %d p_au %p\n", child->p_pid,
1082 // child->p_au);
1083 bcopy(parent->p_au, child->p_au, sizeof(*child->p_au));
1084 /*
1085 * XXXAUDIT: Zero pointers to external memory, or assert they are
1086 * zero?
1087 */
1088}
1089
1090/*
1091 * Free the auditing structure for the process.
1092 */
1093void
1094audit_proc_free(struct proc *p)
1095{
1096
1097 KASSERT(p->p_au != NULL, ("p->p_au == NULL (%d)", p->p_pid));
1098 //printf("audit_proc_free: pid %d p_au %p\n", p->p_pid, p->p_au);
1099 /*
1100 * XXXAUDIT: Assert that external memory pointers are NULL?
1101 */
1102 free(p->p_au, M_AUDITPROC);
1103 p->p_au = NULL;
1104}
| 62#include <bsm/audit_kevents.h>
63
64#include <netinet/in.h>
65#include <netinet/in_pcb.h>
66
67#include <security/audit/audit.h>
68#include <security/audit/audit_private.h>
69
70#include <vm/uma.h>
71
72/*
73 * The AUDIT_EXCESSIVELY_VERBOSE define enables a number of
74 * gratuitously noisy printf's to the console. Due to the
75 * volume, it should be left off unless you want your system
76 * to churn a lot whenever the audit record flow gets high.
77 */
78//#define AUDIT_EXCESSIVELY_VERBOSE
79#ifdef AUDIT_EXCESSIVELY_VERBOSE
80#define AUDIT_PRINTF(x) printf x
81#else
82#define AUDIT_PRINTF(X)
83#endif
84
85static uma_zone_t audit_record_zone;
86static MALLOC_DEFINE(M_AUDITPROC, "audit_proc", "Audit process storage");
87MALLOC_DEFINE(M_AUDITDATA, "audit_data", "Audit data storage");
88MALLOC_DEFINE(M_AUDITPATH, "audit_path", "Audit path storage");
89MALLOC_DEFINE(M_AUDITTEXT, "audit_text", "Audit text storage");
90
91/*
92 * Audit control settings that are set/read by system calls and are
93 * hence non-static.
94 */
95/*
96 * Define the audit control flags.
97 */
98int audit_enabled;
99int audit_suspended;
100
101/*
102 * Flags controlling behavior in low storage situations.
103 * Should we panic if a write fails? Should we fail stop
104 * if we're out of disk space?
105 */
106int audit_panic_on_write_fail;
107int audit_fail_stop;
108
109/*
110 * Are we currently "failing stop" due to out of disk space?
111 */
112static int audit_in_failure;
113
114/*
115 * Global audit statistiscs.
116 */
117struct audit_fstat audit_fstat;
118
119/*
120 * Preselection mask for non-attributable events.
121 */
122struct au_mask audit_nae_mask;
123
124/*
125 * Mutex to protect global variables shared between various threads and
126 * processes.
127 */
128static struct mtx audit_mtx;
129
130/*
131 * Queue of audit records ready for delivery to disk. We insert new
132 * records at the tail, and remove records from the head. Also,
133 * a count of the number of records used for checking queue depth.
134 * In addition, a counter of records that we have allocated but are
135 * not yet in the queue, which is needed to estimate the total
136 * size of the combined set of records outstanding in the system.
137 */
138static TAILQ_HEAD(, kaudit_record) audit_q;
139static int audit_q_len;
140static int audit_pre_q_len;
141
142/*
143 * Audit queue control settings (minimum free, low/high water marks, etc.)
144 */
145struct au_qctrl audit_qctrl;
146
147/*
148 * Condition variable to signal to the worker that it has work to do:
149 * either new records are in the queue, or a log replacement is taking
150 * place.
151 */
152static struct cv audit_cv;
153
154/*
155 * Worker thread that will schedule disk I/O, etc.
156 */
157static struct proc *audit_thread;
158
159/*
160 * When an audit log is rotated, the actual rotation must be performed
161 * by the audit worker thread, as it may have outstanding writes on the
162 * current audit log. audit_replacement_vp holds the vnode replacing
163 * the current vnode. We can't let more than one replacement occur
164 * at a time, so if more than one thread requests a replacement, only
165 * one can have the replacement "in progress" at any given moment. If
166 * a thread tries to replace the audit vnode and discovers a replacement
167 * is already in progress (i.e., audit_replacement_flag != 0), then it
168 * will sleep on audit_replacement_cv waiting its turn to perform a
169 * replacement. When a replacement is completed, this cv is signalled
170 * by the worker thread so a waiting thread can start another replacement.
171 * We also store a credential to perform audit log write operations with.
172 *
173 * The current credential and vnode are thread-local to audit_worker.
174 */
175static struct cv audit_replacement_cv;
176
177static int audit_replacement_flag;
178static struct vnode *audit_replacement_vp;
179static struct ucred *audit_replacement_cred;
180
181/*
182 * Condition variable to signal to the worker that it has work to do:
183 * either new records are in the queue, or a log replacement is taking
184 * place.
185 */
186static struct cv audit_commit_cv;
187
188/*
189 * Condition variable for auditing threads wait on when in fail-stop mode.
190 * Threads wait on this CV forever (and ever), never seeing the light of
191 * day again.
192 */
193static struct cv audit_fail_cv;
194
195/*
196 * Flags related to Kernel->user-space communication.
197 */
198static int audit_file_rotate_wait;
199
200/*
201 * Construct an audit record for the passed thread.
202 */
203static int
204audit_record_ctor(void *mem, int size, void *arg, int flags)
205{
206 struct kaudit_record *ar;
207 struct thread *td;
208
209 KASSERT(sizeof(*ar) == size, ("audit_record_ctor: wrong size"));
210
211 td = arg;
212 ar = mem;
213 bzero(ar, sizeof(*ar));
214 ar->k_ar.ar_magic = AUDIT_RECORD_MAGIC;
215 nanotime(&ar->k_ar.ar_starttime);
216
217 /*
218 * Export the subject credential.
219 *
220 * XXXAUDIT: td_ucred access is OK without proc lock, but some other
221 * fields here may require the proc lock.
222 */
223 cru2x(td->td_ucred, &ar->k_ar.ar_subj_cred);
224 ar->k_ar.ar_subj_ruid = td->td_ucred->cr_ruid;
225 ar->k_ar.ar_subj_rgid = td->td_ucred->cr_rgid;
226 ar->k_ar.ar_subj_egid = td->td_ucred->cr_groups[0];
227 ar->k_ar.ar_subj_auid = td->td_proc->p_au->ai_auid;
228 ar->k_ar.ar_subj_asid = td->td_proc->p_au->ai_asid;
229 ar->k_ar.ar_subj_pid = td->td_proc->p_pid;
230 ar->k_ar.ar_subj_amask = td->td_proc->p_au->ai_mask;
231 ar->k_ar.ar_subj_term = td->td_proc->p_au->ai_termid;
232 bcopy(td->td_proc->p_comm, ar->k_ar.ar_subj_comm, MAXCOMLEN);
233
234 return (0);
235}
236
237static void
238audit_record_dtor(void *mem, int size, void *arg)
239{
240 struct kaudit_record *ar;
241
242 KASSERT(sizeof(*ar) == size, ("audit_record_dtor: wrong size"));
243
244 ar = mem;
245 if (ar->k_ar.ar_arg_upath1 != NULL)
246 free(ar->k_ar.ar_arg_upath1, M_AUDITPATH);
247 if (ar->k_ar.ar_arg_upath2 != NULL)
248 free(ar->k_ar.ar_arg_upath2, M_AUDITPATH);
249 if (ar->k_ar.ar_arg_text != NULL)
250 free(ar->k_ar.ar_arg_text, M_AUDITTEXT);
251 if (ar->k_udata != NULL)
252 free(ar->k_udata, M_AUDITDATA);
253}
254
255/*
256 * XXXAUDIT: Should adjust comments below to make it clear that we get to
257 * this point only if we believe we have storage, so not having space here
258 * is a violation of invariants derived from administrative procedures.
259 * I.e., someone else has written to the audit partition, leaving less space
260 * than we accounted for.
261 */
262static int
263audit_record_write(struct vnode *vp, struct kaudit_record *ar,
264 struct ucred *cred, struct thread *td)
265{
266 int ret;
267 long temp;
268 struct au_record *bsm;
269 struct vattr vattr;
270 struct statfs *mnt_stat = &vp->v_mount->mnt_stat;
271 int vfslocked;
272
273 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
274
275 /*
276 * First, gather statistics on the audit log file and file system
277 * so that we know how we're doing on space. In both cases,
278 * if we're unable to perform the operation, we drop the record
279 * and return. However, this is arguably an assertion failure.
280 * XXX Need a FreeBSD equivalent.
281 */
282 ret = VFS_STATFS(vp->v_mount, mnt_stat, td);
283 if (ret)
284 goto out;
285
286 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
287 ret = VOP_GETATTR(vp, &vattr, cred, td);
288 VOP_UNLOCK(vp, 0, td);
289 if (ret)
290 goto out;
291
292 /* update the global stats struct */
293 audit_fstat.af_currsz = vattr.va_size;
294
295 /*
296 * XXX Need to decide what to do if the trigger to the audit daemon
297 * fails.
298 */
299
300 /*
301 * If we fall below minimum free blocks (hard limit), tell the audit
302 * daemon to force a rotation off of the file system. We also stop
303 * writing, which means this audit record is probably lost.
304 * If we fall below the minimum percent free blocks (soft limit),
305 * then kindly suggest to the audit daemon to do something.
306 */
307 if (mnt_stat->f_bfree < AUDIT_HARD_LIMIT_FREE_BLOCKS) {
308 send_trigger(AUDIT_TRIGGER_NO_SPACE);
309 /* Hopefully userspace did something about all the previous
310 * triggers that were sent prior to this critical condition.
311 * If fail-stop is set, then we're done; goodnight Gracie.
312 */
313 if (audit_fail_stop)
314 panic("Audit log space exhausted and fail-stop set.");
315 else {
316 audit_suspended = 1;
317 ret = ENOSPC;
318 goto out;
319 }
320 } else
321 /*
322 * Send a message to the audit daemon that disk space
323 * is getting low.
324 *
325 * XXXAUDIT: Check math and block size calculation here.
326 */
327 if (audit_qctrl.aq_minfree != 0) {
328 temp = mnt_stat->f_blocks / (100 /
329 audit_qctrl.aq_minfree);
330 if (mnt_stat->f_bfree < temp)
331 send_trigger(AUDIT_TRIGGER_LOW_SPACE);
332 }
333
334 /* Check if the current log file is full; if so, call for
335 * a log rotate. This is not an exact comparison; we may
336 * write some records over the limit. If that's not
337 * acceptable, then add a fudge factor here.
338 */
339 if ((audit_fstat.af_filesz != 0) &&
340 (audit_file_rotate_wait == 0) &&
341 (vattr.va_size >= audit_fstat.af_filesz)) {
342 audit_file_rotate_wait = 1;
343 send_trigger(AUDIT_TRIGGER_OPEN_NEW);
344 }
345
346 /*
347 * If the estimated amount of audit data in the audit event queue
348 * (plus records allocated but not yet queued) has reached the
349 * amount of free space on the disk, then we need to go into an
350 * audit fail stop state, in which we do not permit the
351 * allocation/committing of any new audit records. We continue to
352 * process packets but don't allow any activities that might
353 * generate new records. In the future, we might want to detect
354 * when space is available again and allow operation to continue,
355 * but this behavior is sufficient to meet fail stop requirements
356 * in CAPP.
357 */
358 if (audit_fail_stop &&
359 (unsigned long)
360 ((audit_q_len + audit_pre_q_len + 1) * MAX_AUDIT_RECORD_SIZE) /
361 mnt_stat->f_bsize >= (unsigned long)(mnt_stat->f_bfree)) {
362 printf(
363 "audit_worker: free space below size of audit queue, failing stop\n");
364 audit_in_failure = 1;
365 }
366
367 /*
368 * If there is a user audit record attached to the kernel record,
369 * then write the user record.
370 */
371 /* XXX Need to decide a few things here: IF the user audit
372 * record is written, but the write of the kernel record fails,
373 * what to do? Should the kernel record come before or after the
374 * user record? For now, we write the user record first, and
375 * we ignore errors.
376 */
377 if (ar->k_ar_commit & AR_COMMIT_USER) {
378 /*
379 * Try submitting the record to any active audit pipes.
380 */
381 audit_pipe_submit((void *)ar->k_udata, ar->k_ulen);
382
383 /*
384 * And to disk.
385 */
386 ret = vn_rdwr(UIO_WRITE, vp, (void *)ar->k_udata, ar->k_ulen,
387 (off_t)0, UIO_SYSSPACE, IO_APPEND|IO_UNIT, cred, NULL,
388 NULL, td);
389 if (ret)
390 goto out;
391 }
392
393 /*
394 * Convert the internal kernel record to BSM format and write it
395 * out if everything's OK.
396 */
397 if (!(ar->k_ar_commit & AR_COMMIT_KERNEL)) {
398 ret = 0;
399 goto out;
400 }
401
402 /*
403 * XXXAUDIT: Should we actually allow this conversion to fail? With
404 * sleeping memory allocation and invariants checks, perhaps not.
405 */
406 ret = kaudit_to_bsm(ar, &bsm);
407 if (ret == BSM_NOAUDIT) {
408 ret = 0;
409 goto out;
410 }
411
412 /*
413 * XXX: We drop the record on BSM conversion failure, but really
414 * this is an assertion failure.
415 */
416 if (ret == BSM_FAILURE) {
417 AUDIT_PRINTF(("BSM conversion failure\n"));
418 ret = EINVAL;
419 goto out;
420 }
421
422 /*
423 * Try submitting the record to any active audit pipes.
424 */
425 audit_pipe_submit((void *)bsm->data, bsm->len);
426
427 /*
428 * XXX
429 * We should break the write functionality away from the BSM record
430 * generation and have the BSM generation done before this function
431 * is called. This function will then take the BSM record as a
432 * parameter.
433 */
434 ret = (vn_rdwr(UIO_WRITE, vp, (void *)bsm->data, bsm->len,
435 (off_t)0, UIO_SYSSPACE, IO_APPEND|IO_UNIT, cred, NULL, NULL, td));
436
437 kau_free(bsm);
438
439out:
440 /*
441 * When we're done processing the current record, we have to
442 * check to see if we're in a failure mode, and if so, whether
443 * this was the last record left to be drained. If we're done
444 * draining, then we fsync the vnode and panic.
445 */
446 if (audit_in_failure &&
447 audit_q_len == 0 && audit_pre_q_len == 0) {
448 VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td);
449 (void)VOP_FSYNC(vp, MNT_WAIT, td);
450 VOP_UNLOCK(vp, 0, td);
451 panic("Audit store overflow; record queue drained.");
452 }
453
454 VFS_UNLOCK_GIANT(vfslocked);
455
456 return (ret);
457}
458
459/*
460 * The audit_worker thread is responsible for watching the event queue,
461 * dequeueing records, converting them to BSM format, and committing them to
462 * disk. In order to minimize lock thrashing, records are dequeued in sets
463 * to a thread-local work queue. In addition, the audit_work performs the
464 * actual exchange of audit log vnode pointer, as audit_vp is a thread-local
465 * variable.
466 */
467static void
468audit_worker(void *arg)
469{
470 int do_replacement_signal, error;
471 TAILQ_HEAD(, kaudit_record) ar_worklist;
472 struct kaudit_record *ar;
473 struct vnode *audit_vp, *old_vp;
474 int vfslocked;
475
476 struct ucred *audit_cred, *old_cred;
477 struct thread *audit_td;
478
479 AUDIT_PRINTF(("audit_worker starting\n"));
480
481 /*
482 * These are thread-local variables requiring no synchronization.
483 */
484 TAILQ_INIT(&ar_worklist);
485 audit_cred = NULL;
486 audit_td = curthread;
487 audit_vp = NULL;
488
489 mtx_lock(&audit_mtx);
490 while (1) {
491 /*
492 * First priority: replace the audit log target if requested.
493 * Accessing the vnode here requires dropping the audit_mtx;
494 * in case another replacement was scheduled while the mutex
495 * was released, we loop.
496 *
497 * XXX It could well be we should drain existing records
498 * first to ensure that the timestamps and ordering
499 * are right.
500 */
501 do_replacement_signal = 0;
502 while (audit_replacement_flag != 0) {
503 old_cred = audit_cred;
504 old_vp = audit_vp;
505 audit_cred = audit_replacement_cred;
506 audit_vp = audit_replacement_vp;
507 audit_replacement_cred = NULL;
508 audit_replacement_vp = NULL;
509 audit_replacement_flag = 0;
510
511 audit_enabled = (audit_vp != NULL);
512
513 /*
514 * XXX: What to do about write failures here?
515 */
516 if (old_vp != NULL) {
517 AUDIT_PRINTF(("Closing old audit file\n"));
518 mtx_unlock(&audit_mtx);
519 vfslocked = VFS_LOCK_GIANT(old_vp->v_mount);
520 vn_close(old_vp, AUDIT_CLOSE_FLAGS, old_cred,
521 audit_td);
522 VFS_UNLOCK_GIANT(vfslocked);
523 crfree(old_cred);
524 mtx_lock(&audit_mtx);
525 old_cred = NULL;
526 old_vp = NULL;
527 AUDIT_PRINTF(("Audit file closed\n"));
528 }
529 if (audit_vp != NULL) {
530 AUDIT_PRINTF(("Opening new audit file\n"));
531 }
532 do_replacement_signal = 1;
533 }
534 /*
535 * Signal that replacement have occurred to wake up and
536 * start any other replacements started in parallel. We can
537 * continue about our business in the mean time. We
538 * broadcast so that both new replacements can be inserted,
539 * but also so that the source(s) of replacement can return
540 * successfully.
541 */
542 if (do_replacement_signal)
543 cv_broadcast(&audit_replacement_cv);
544
545 /*
546 * Next, check to see if we have any records to drain into
547 * the vnode. If not, go back to waiting for an event.
548 */
549 if (TAILQ_EMPTY(&audit_q)) {
550 AUDIT_PRINTF(("audit_worker waiting\n"));
551 cv_wait(&audit_cv, &audit_mtx);
552 AUDIT_PRINTF(("audit_worker woken up\n"));
553 AUDIT_PRINTF(("audit_worker: new vp = %p; value of flag %d\n",
554 audit_replacement_vp, audit_replacement_flag));
555 continue;
556 }
557
558 /*
559 * If we have records, but there's no active vnode to write
560 * to, drain the record queue. Generally, we prevent the
561 * unnecessary allocation of records elsewhere, but we need
562 * to allow for races between conditional allocation and
563 * queueing. Go back to waiting when we're done.
564 */
565 if (audit_vp == NULL) {
566 while ((ar = TAILQ_FIRST(&audit_q))) {
567 TAILQ_REMOVE(&audit_q, ar, k_q);
568 uma_zfree(audit_record_zone, ar);
569 audit_q_len--;
570 /*
571 * XXXRW: Why broadcast if we hold the
572 * mutex and know that audit_vp is NULL?
573 */
574 if (audit_q_len <= audit_qctrl.aq_lowater)
575 cv_broadcast(&audit_commit_cv);
576 }
577 continue;
578 }
579
580 /*
581 * We have both records to write and an active vnode to write
582 * to. Dequeue a record, and start the write. Eventually,
583 * it might make sense to dequeue several records and perform
584 * our own clustering, if the lower layers aren't doing it
585 * automatically enough.
586 */
587 while ((ar = TAILQ_FIRST(&audit_q))) {
588 TAILQ_REMOVE(&audit_q, ar, k_q);
589 audit_q_len--;
590 if (audit_q_len <= audit_qctrl.aq_lowater)
591 cv_broadcast(&audit_commit_cv);
592 TAILQ_INSERT_TAIL(&ar_worklist, ar, k_q);
593 }
594
595 mtx_unlock(&audit_mtx);
596 while ((ar = TAILQ_FIRST(&ar_worklist))) {
597 TAILQ_REMOVE(&ar_worklist, ar, k_q);
598 if (audit_vp != NULL) {
599 error = audit_record_write(audit_vp, ar,
600 audit_cred, audit_td);
601 if (error && audit_panic_on_write_fail)
602 panic("audit_worker: write error %d\n",
603 error);
604 else if (error)
605 printf("audit_worker: write error %d\n",
606 error);
607 }
608 uma_zfree(audit_record_zone, ar);
609 }
610 mtx_lock(&audit_mtx);
611 }
612}
613
614/*
615 * Initialize the Audit subsystem: configuration state, work queue,
616 * synchronization primitives, worker thread, and trigger device node. Also
617 * call into the BSM assembly code to initialize it.
618 */
619static void
620audit_init(void)
621{
622 int error;
623
624 printf("Security auditing service present\n");
625 audit_enabled = 0;
626 audit_suspended = 0;
627 audit_panic_on_write_fail = 0;
628 audit_fail_stop = 0;
629 audit_in_failure = 0;
630
631 audit_replacement_vp = NULL;
632 audit_replacement_cred = NULL;
633 audit_replacement_flag = 0;
634
635 audit_fstat.af_filesz = 0; /* '0' means unset, unbounded */
636 audit_fstat.af_currsz = 0;
637 audit_nae_mask.am_success = AU_NULL;
638 audit_nae_mask.am_failure = AU_NULL;
639
640 TAILQ_INIT(&audit_q);
641 audit_q_len = 0;
642 audit_pre_q_len = 0;
643 audit_qctrl.aq_hiwater = AQ_HIWATER;
644 audit_qctrl.aq_lowater = AQ_LOWATER;
645 audit_qctrl.aq_bufsz = AQ_BUFSZ;
646 audit_qctrl.aq_minfree = AU_FS_MINFREE;
647
648 mtx_init(&audit_mtx, "audit_mtx", NULL, MTX_DEF);
649 cv_init(&audit_cv, "audit_cv");
650 cv_init(&audit_replacement_cv, "audit_replacement_cv");
651 cv_init(&audit_commit_cv, "audit_commit_cv");
652 cv_init(&audit_fail_cv, "audit_fail_cv");
653
654 audit_record_zone = uma_zcreate("audit_record_zone",
655 sizeof(struct kaudit_record), audit_record_ctor,
656 audit_record_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
657
658 /* Initialize the BSM audit subsystem. */
659 kau_init();
660
661 audit_file_rotate_wait = 0;
662 audit_trigger_init();
663
664 /* Register shutdown handler. */
665 EVENTHANDLER_REGISTER(shutdown_pre_sync, audit_shutdown, NULL,
666 SHUTDOWN_PRI_FIRST);
667
668 error = kthread_create(audit_worker, NULL, &audit_thread, RFHIGHPID,
669 0, "audit_worker");
670 if (error != 0)
671 panic("audit_init: kthread_create returned %d", error);
672}
673
674SYSINIT(audit_init, SI_SUB_AUDIT, SI_ORDER_FIRST, audit_init, NULL)
675
676/*
677 * audit_rotate_vnode() is called by a user or kernel thread to configure or
678 * de-configure auditing on a vnode. The arguments are the replacement
679 * credential and vnode to substitute for the current credential and vnode,
680 * if any. If either is set to NULL, both should be NULL, and this is used
681 * to indicate that audit is being disabled. The real work is done in the
682 * audit_worker thread, but audit_rotate_vnode() waits synchronously for that
683 * to complete.
684 *
685 * The vnode should be referenced and opened by the caller. The credential
686 * should be referenced. audit_rotate_vnode() will own both references as of
687 * this call, so the caller should not release either.
688 *
689 * XXXAUDIT: Review synchronize communication logic. Really, this is a
690 * message queue of depth 1.
691 *
692 * XXXAUDIT: Enhance the comments below to indicate that we are basically
693 * acquiring ownership of the communications queue, inserting our message,
694 * and waiting for an acknowledgement.
695 */
696void
697audit_rotate_vnode(struct ucred *cred, struct vnode *vp)
698{
699
700 /*
701 * If other parallel log replacements have been requested, we wait
702 * until they've finished before continuing.
703 */
704 mtx_lock(&audit_mtx);
705 while (audit_replacement_flag != 0) {
706 AUDIT_PRINTF(("audit_rotate_vnode: sleeping to wait for "
707 "flag\n"));
708 cv_wait(&audit_replacement_cv, &audit_mtx);
709 AUDIT_PRINTF(("audit_rotate_vnode: woken up (flag %d)\n",
710 audit_replacement_flag));
711 }
712 audit_replacement_cred = cred;
713 audit_replacement_flag = 1;
714 audit_replacement_vp = vp;
715
716 /*
717 * Wake up the audit worker to perform the exchange once we
718 * release the mutex.
719 */
720 cv_signal(&audit_cv);
721
722 /*
723 * Wait for the audit_worker to broadcast that a replacement has
724 * taken place; we know that once this has happened, our vnode
725 * has been replaced in, so we can return successfully.
726 */
727 AUDIT_PRINTF(("audit_rotate_vnode: waiting for news of "
728 "replacement\n"));
729 cv_wait(&audit_replacement_cv, &audit_mtx);
730 AUDIT_PRINTF(("audit_rotate_vnode: change acknowledged by "
731 "audit_worker (flag " "now %d)\n", audit_replacement_flag));
732 mtx_unlock(&audit_mtx);
733
734 audit_file_rotate_wait = 0; /* We can now request another rotation */
735}
736
737/*
738 * Drain the audit queue and close the log at shutdown. Note that this can
739 * be called both from the system shutdown path and also from audit
740 * configuration syscalls, so 'arg' and 'howto' are ignored.
741 */
742void
743audit_shutdown(void *arg, int howto)
744{
745
746 audit_rotate_vnode(NULL, NULL);
747}
748
749/*
750 * Return the current thread's audit record, if any.
751 */
752__inline__ struct kaudit_record *
753currecord(void)
754{
755
756 return (curthread->td_ar);
757}
758
759/*
760 * MPSAFE
761 *
762 * XXXAUDIT: There are a number of races present in the code below due to
763 * release and re-grab of the mutex. The code should be revised to become
764 * slightly less racy.
765 *
766 * XXXAUDIT: Shouldn't there be logic here to sleep waiting on available
767 * pre_q space, suspending the system call until there is room?
768 */
769struct kaudit_record *
770audit_new(int event, struct thread *td)
771{
772 struct kaudit_record *ar;
773 int no_record;
774
775 mtx_lock(&audit_mtx);
776 no_record = (audit_suspended || !audit_enabled);
777 mtx_unlock(&audit_mtx);
778 if (no_record)
779 return (NULL);
780
781 /*
782 * XXX: The number of outstanding uncommitted audit records is
783 * limited to the number of concurrent threads servicing system
784 * calls in the kernel.
785 */
786 ar = uma_zalloc_arg(audit_record_zone, td, M_WAITOK);
787 ar->k_ar.ar_event = event;
788
789 mtx_lock(&audit_mtx);
790 audit_pre_q_len++;
791 mtx_unlock(&audit_mtx);
792
793 return (ar);
794}
795
796/*
797 * MPSAFE
798 */
799void
800audit_commit(struct kaudit_record *ar, int error, int retval)
801{
802 int sorf;
803 struct au_mask *aumask;
804
805 if (ar == NULL)
806 return;
807
808 /*
809 * Decide whether to commit the audit record by checking the
810 * error value from the system call and using the appropriate
811 * audit mask.
812 *
813 * XXXAUDIT: Synchronize access to audit_nae_mask?
814 */
815 if (ar->k_ar.ar_subj_auid == AU_DEFAUDITID)
816 aumask = &audit_nae_mask;
817 else
818 aumask = &ar->k_ar.ar_subj_amask;
819
820 if (error)
821 sorf = AU_PRS_FAILURE;
822 else
823 sorf = AU_PRS_SUCCESS;
824
825 switch(ar->k_ar.ar_event) {
826
827 case AUE_OPEN_RWTC:
828 /* The open syscall always writes a AUE_OPEN_RWTC event; change
829 * it to the proper type of event based on the flags and the
830 * error value.
831 */
832 ar->k_ar.ar_event = flags_and_error_to_openevent(
833 ar->k_ar.ar_arg_fflags, error);
834 break;
835
836 case AUE_SYSCTL:
837 ar->k_ar.ar_event = ctlname_to_sysctlevent(
838 ar->k_ar.ar_arg_ctlname, ar->k_ar.ar_valid_arg);
839 break;
840
841 case AUE_AUDITON:
842 /* Convert the auditon() command to an event */
843 ar->k_ar.ar_event = auditon_command_event(ar->k_ar.ar_arg_cmd);
844 break;
845 }
846
847 if (au_preselect(ar->k_ar.ar_event, aumask, sorf) != 0)
848 ar->k_ar_commit |= AR_COMMIT_KERNEL;
849
850 /*
851 * XXXRW: Why is this necessary? Should we ever accept a record that
852 * we're not willing to commit?
853 */
854 if ((ar->k_ar_commit & (AR_COMMIT_USER | AR_COMMIT_KERNEL)) == 0) {
855 mtx_lock(&audit_mtx);
856 audit_pre_q_len--;
857 mtx_unlock(&audit_mtx);
858 uma_zfree(audit_record_zone, ar);
859 return;
860 }
861
862 ar->k_ar.ar_errno = error;
863 ar->k_ar.ar_retval = retval;
864
865 /*
866 * We might want to do some system-wide post-filtering
867 * here at some point.
868 */
869
870 /*
871 * Timestamp system call end.
872 */
873 nanotime(&ar->k_ar.ar_endtime);
874
875 mtx_lock(&audit_mtx);
876
877 /*
878 * Note: it could be that some records initiated while audit was
879 * enabled should still be committed?
880 */
881 if (audit_suspended || !audit_enabled) {
882 audit_pre_q_len--;
883 mtx_unlock(&audit_mtx);
884 uma_zfree(audit_record_zone, ar);
885 return;
886 }
887
888 /*
889 * Constrain the number of committed audit records based on
890 * the configurable parameter.
891 */
892 while (audit_q_len >= audit_qctrl.aq_hiwater) {
893 AUDIT_PRINTF(("audit_commit: sleeping to wait for "
894 "audit queue to drain below high water mark\n"));
895 cv_wait(&audit_commit_cv, &audit_mtx);
896 AUDIT_PRINTF(("audit_commit: woke up waiting for "
897 "audit queue draining\n"));
898 }
899
900 TAILQ_INSERT_TAIL(&audit_q, ar, k_q);
901 audit_q_len++;
902 audit_pre_q_len--;
903 cv_signal(&audit_cv);
904 mtx_unlock(&audit_mtx);
905}
906
907/*
908 * audit_syscall_enter() is called on entry to each system call. It is
909 * responsible for deciding whether or not to audit the call (preselection),
910 * and if so, allocating a per-thread audit record. audit_new() will fill in
911 * basic thread/credential properties.
912 */
913void
914audit_syscall_enter(unsigned short code, struct thread *td)
915{
916 int audit_event;
917 struct au_mask *aumask;
918
919 KASSERT(td->td_ar == NULL, ("audit_syscall_enter: td->td_ar != NULL"));
920
921 /*
922 * In FreeBSD, each ABI has its own system call table, and hence
923 * mapping of system call codes to audit events. Convert the code to
924 * an audit event identifier using the process system call table
925 * reference. In Darwin, there's only one, so we use the global
926 * symbol for the system call table.
927 *
928 * XXXAUDIT: Should we audit that a bad system call was made, and if
929 * so, how?
930 */
931 if (code >= td->td_proc->p_sysent->sv_size)
932 return;
933
934 audit_event = td->td_proc->p_sysent->sv_table[code].sy_auevent;
935 if (audit_event == AUE_NULL)
936 return;
937
938 /*
939 * Check which audit mask to use; either the kernel non-attributable
940 * event mask or the process audit mask.
941 */
942 if (td->td_proc->p_au->ai_auid == AU_DEFAUDITID)
943 aumask = &audit_nae_mask;
944 else
945 aumask = &td->td_proc->p_au->ai_mask;
946
947 /*
948 * Allocate an audit record, if preselection allows it, and store
949 * in the thread for later use.
950 */
951 if (au_preselect(audit_event, aumask,
952 AU_PRS_FAILURE | AU_PRS_SUCCESS)) {
953 /*
954 * If we're out of space and need to suspend unprivileged
955 * processes, do that here rather than trying to allocate
956 * another audit record.
957 *
958 * XXXRW: We might wish to be able to continue here in the
959 * future, if the system recovers. That should be possible
960 * by means of checking the condition in a loop around
961 * cv_wait(). It might be desirable to reevaluate whether an
962 * audit record is still required for this event by
963 * re-calling au_preselect().
964 */
965 if (audit_in_failure && suser(td) != 0) {
966 cv_wait(&audit_fail_cv, &audit_mtx);
967 panic("audit_failing_stop: thread continued");
968 }
969 td->td_ar = audit_new(audit_event, td);
970 } else
971 td->td_ar = NULL;
972}
973
974/*
975 * audit_syscall_exit() is called from the return of every system call, or in
976 * the event of exit1(), during the execution of exit1(). It is responsible
977 * for committing the audit record, if any, along with return condition.
978 */
979void
980audit_syscall_exit(int error, struct thread *td)
981{
982 int retval;
983
984 /*
985 * Commit the audit record as desired; once we pass the record
986 * into audit_commit(), the memory is owned by the audit
987 * subsystem.
988 * The return value from the system call is stored on the user
989 * thread. If there was an error, the return value is set to -1,
990 * imitating the behavior of the cerror routine.
991 */
992 if (error)
993 retval = -1;
994 else
995 retval = td->td_retval[0];
996
997 audit_commit(td->td_ar, error, retval);
998 if (td->td_ar != NULL)
999 AUDIT_PRINTF(("audit record committed by pid %d\n",
1000 td->td_proc->p_pid));
1001 td->td_ar = NULL;
1002
1003}
1004
1005/*
1006 * Allocate storage for a new process (init, or otherwise).
1007 */
1008void
1009audit_proc_alloc(struct proc *p)
1010{
1011
1012 KASSERT(p->p_au == NULL, ("audit_proc_alloc: p->p_au != NULL (%d)",
1013 p->p_pid));
1014 p->p_au = malloc(sizeof(*(p->p_au)), M_AUDITPROC, M_WAITOK);
1015 /* XXXAUDIT: Zero? Slab allocate? */
1016 //printf("audit_proc_alloc: pid %d p_au %p\n", p->p_pid, p->p_au);
1017}
1018
1019/*
1020 * Allocate storage for a new thread.
1021 */
1022void
1023audit_thread_alloc(struct thread *td)
1024{
1025
1026 td->td_ar = NULL;
1027}
1028
1029/*
1030 * Thread destruction.
1031 */
1032void
1033audit_thread_free(struct thread *td)
1034{
1035
1036 KASSERT(td->td_ar == NULL, ("audit_thread_free: td_ar != NULL"));
1037}
1038
1039/*
1040 * Initialize the audit information for the a process, presumably the first
1041 * process in the system.
1042 * XXX It is not clear what the initial values should be for audit ID,
1043 * session ID, etc.
1044 */
1045void
1046audit_proc_kproc0(struct proc *p)
1047{
1048
1049 KASSERT(p->p_au != NULL, ("audit_proc_kproc0: p->p_au == NULL (%d)",
1050 p->p_pid));
1051 //printf("audit_proc_kproc0: pid %d p_au %p\n", p->p_pid, p->p_au);
1052 bzero(p->p_au, sizeof(*(p)->p_au));
1053}
1054
1055void
1056audit_proc_init(struct proc *p)
1057{
1058
1059 KASSERT(p->p_au != NULL, ("audit_proc_init: p->p_au == NULL (%d)",
1060 p->p_pid));
1061 //printf("audit_proc_init: pid %d p_au %p\n", p->p_pid, p->p_au);
1062 bzero(p->p_au, sizeof(*(p)->p_au));
1063 p->p_au->ai_auid = AU_DEFAUDITID;
1064}
1065
1066/*
1067 * Copy the audit info from the parent process to the child process when
1068 * a fork takes place.
1069 */
1070void
1071audit_proc_fork(struct proc *parent, struct proc *child)
1072{
1073
1074 PROC_LOCK_ASSERT(parent, MA_OWNED);
1075 PROC_LOCK_ASSERT(child, MA_OWNED);
1076 KASSERT(parent->p_au != NULL,
1077 ("audit_proc_fork: parent->p_au == NULL (%d)", parent->p_pid));
1078 KASSERT(child->p_au != NULL,
1079 ("audit_proc_fork: child->p_au == NULL (%d)", child->p_pid));
1080 //printf("audit_proc_fork: parent pid %d p_au %p\n", parent->p_pid,
1081 // parent->p_au);
1082 //printf("audit_proc_fork: child pid %d p_au %p\n", child->p_pid,
1083 // child->p_au);
1084 bcopy(parent->p_au, child->p_au, sizeof(*child->p_au));
1085 /*
1086 * XXXAUDIT: Zero pointers to external memory, or assert they are
1087 * zero?
1088 */
1089}
1090
1091/*
1092 * Free the auditing structure for the process.
1093 */
1094void
1095audit_proc_free(struct proc *p)
1096{
1097
1098 KASSERT(p->p_au != NULL, ("p->p_au == NULL (%d)", p->p_pid));
1099 //printf("audit_proc_free: pid %d p_au %p\n", p->p_pid, p->p_au);
1100 /*
1101 * XXXAUDIT: Assert that external memory pointers are NULL?
1102 */
1103 free(p->p_au, M_AUDITPROC);
1104 p->p_au = NULL;
1105}
|