1 /* 2 * Copyright (c) 1999-2005 Apple Computer, Inc. 3 * Copyright (c) 2006 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/sysproto.h> 52 #include <sys/sysent.h> 53 #include <sys/systm.h> 54 #include <sys/ucred.h> 55 #include <sys/uio.h> 56 #include <sys/un.h> 57 #include <sys/unistd.h> 58 #include <sys/vnode.h> 59 60 #include <bsm/audit.h> 61 #include <bsm/audit_internal.h> 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 * Worker thread that will schedule disk I/O, etc. 74 */ 75 static struct proc *audit_thread; 76 77 /* 78 * When an audit log is rotated, the actual rotation must be performed by the 79 * audit worker thread, as it may have outstanding writes on the current 80 * audit log. audit_replacement_vp holds the vnode replacing the current 81 * vnode. We can't let more than one replacement occur at a time, so if more 82 * than one thread requests a replacement, only one can have the replacement 83 * "in progress" at any given moment. If a thread tries to replace the audit 84 * vnode and discovers a replacement is already in progress (i.e., 85 * audit_replacement_flag != 0), then it will sleep on audit_replacement_cv 86 * waiting its turn to perform a replacement. When a replacement is 87 * completed, this cv is signalled by the worker thread so a waiting thread 88 * can start another replacement. We also store a credential to perform 89 * audit log write operations with. 90 * 91 * The current credential and vnode are thread-local to audit_worker. 92 */ 93 static struct cv audit_replacement_cv; 94 95 static int audit_replacement_flag; 96 static struct vnode *audit_replacement_vp; 97 static struct ucred *audit_replacement_cred; 98 99 /* 100 * Flags related to Kernel->user-space communication. 101 */ 102 static int audit_file_rotate_wait; 103 104 /* 105 * XXXAUDIT: Should adjust comments below to make it clear that we get to 106 * this point only if we believe we have storage, so not having space here is 107 * a violation of invariants derived from administrative procedures. I.e., 108 * someone else has written to the audit partition, leaving less space than 109 * we accounted for. 110 */ 111 static int 112 audit_record_write(struct vnode *vp, struct ucred *cred, struct thread *td, 113 void *data, size_t len) 114 { 115 int ret; 116 long temp; 117 struct vattr vattr; 118 struct statfs *mnt_stat; 119 int vfslocked; 120 121 if (vp == NULL) 122 return (0); 123 124 mnt_stat = &vp->v_mount->mnt_stat; 125 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 126 127 /* 128 * First, gather statistics on the audit log file and file system so 129 * that we know how we're doing on space. In both cases, if we're 130 * unable to perform the operation, we drop the record and return. 131 * However, this is arguably an assertion failure. 132 * XXX Need a FreeBSD equivalent. 133 */ 134 ret = VFS_STATFS(vp->v_mount, mnt_stat, td); 135 if (ret) 136 goto out; 137 138 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td); 139 ret = VOP_GETATTR(vp, &vattr, cred, td); 140 VOP_UNLOCK(vp, 0, td); 141 if (ret) 142 goto out; 143 144 /* update the global stats struct */ 145 audit_fstat.af_currsz = vattr.va_size; 146 147 /* 148 * XXX Need to decide what to do if the trigger to the audit daemon 149 * fails. 150 */ 151 152 /* 153 * If we fall below minimum free blocks (hard limit), tell the audit 154 * daemon to force a rotation off of the file system. We also stop 155 * writing, which means this audit record is probably lost. If we 156 * fall below the minimum percent free blocks (soft limit), then 157 * kindly suggest to the audit daemon to do something. 158 */ 159 if (mnt_stat->f_bfree < AUDIT_HARD_LIMIT_FREE_BLOCKS) { 160 (void)send_trigger(AUDIT_TRIGGER_NO_SPACE); 161 /* 162 * Hopefully userspace did something about all the previous 163 * triggers that were sent prior to this critical condition. 164 * If fail-stop is set, then we're done; goodnight Gracie. 165 */ 166 if (audit_fail_stop) 167 panic("Audit log space exhausted and fail-stop set."); 168 else { 169 audit_suspended = 1; 170 ret = ENOSPC; 171 goto out; 172 } 173 } else 174 /* 175 * Send a message to the audit daemon that disk space is 176 * getting low. 177 * 178 * XXXAUDIT: Check math and block size calculation here. 179 */ 180 if (audit_qctrl.aq_minfree != 0) { 181 temp = mnt_stat->f_blocks / (100 / 182 audit_qctrl.aq_minfree); 183 if (mnt_stat->f_bfree < temp) 184 (void)send_trigger(AUDIT_TRIGGER_LOW_SPACE); 185 } 186 187 /* 188 * Check if the current log file is full; if so, call for a log 189 * rotate. This is not an exact comparison; we may write some records 190 * over the limit. If that's not acceptable, then add a fudge factor 191 * here. 192 */ 193 if ((audit_fstat.af_filesz != 0) && 194 (audit_file_rotate_wait == 0) && 195 (vattr.va_size >= audit_fstat.af_filesz)) { 196 audit_file_rotate_wait = 1; 197 (void)send_trigger(AUDIT_TRIGGER_OPEN_NEW); 198 } 199 200 /* 201 * If the estimated amount of audit data in the audit event queue 202 * (plus records allocated but not yet queued) has reached the amount 203 * of free space on the disk, then we need to go into an audit fail 204 * stop state, in which we do not permit the allocation/committing of 205 * any new audit records. We continue to process packets but don't 206 * allow any activities that might generate new records. In the 207 * future, we might want to detect when space is available again and 208 * allow operation to continue, but this behavior is sufficient to 209 * meet fail stop requirements in CAPP. 210 */ 211 if (audit_fail_stop && 212 (unsigned long) 213 ((audit_q_len + audit_pre_q_len + 1) * MAX_AUDIT_RECORD_SIZE) / 214 mnt_stat->f_bsize >= (unsigned long)(mnt_stat->f_bfree)) { 215 printf("audit_record_write: free space below size of audit " 216 "queue, failing stop\n"); 217 audit_in_failure = 1; 218 } 219 220 ret = vn_rdwr(UIO_WRITE, vp, data, len, (off_t)0, UIO_SYSSPACE, 221 IO_APPEND|IO_UNIT, cred, NULL, NULL, td); 222 223 out: 224 /* 225 * When we're done processing the current record, we have to check to 226 * see if we're in a failure mode, and if so, whether this was the 227 * last record left to be drained. If we're done draining, then we 228 * fsync the vnode and panic. 229 */ 230 if (audit_in_failure && audit_q_len == 0 && audit_pre_q_len == 0) { 231 VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td); 232 (void)VOP_FSYNC(vp, MNT_WAIT, td); 233 VOP_UNLOCK(vp, 0, td); 234 panic("Audit store overflow; record queue drained."); 235 } 236 237 VFS_UNLOCK_GIANT(vfslocked); 238 239 return (ret); 240 } 241 242 /* 243 * If an appropriate signal has been received rotate the audit log based on 244 * the global replacement variables. Signal consumers as needed that the 245 * rotation has taken place. 246 * 247 * XXXRW: The global variables and CVs used to signal the audit_worker to 248 * perform a rotation are essentially a message queue of depth 1. It would 249 * be much nicer to actually use a message queue. 250 */ 251 static void 252 audit_worker_rotate(struct ucred **audit_credp, struct vnode **audit_vpp, 253 struct thread *audit_td) 254 { 255 int do_replacement_signal, vfslocked; 256 struct ucred *old_cred; 257 struct vnode *old_vp; 258 259 mtx_assert(&audit_mtx, MA_OWNED); 260 261 do_replacement_signal = 0; 262 while (audit_replacement_flag != 0) { 263 old_cred = *audit_credp; 264 old_vp = *audit_vpp; 265 *audit_credp = audit_replacement_cred; 266 *audit_vpp = audit_replacement_vp; 267 audit_replacement_cred = NULL; 268 audit_replacement_vp = NULL; 269 audit_replacement_flag = 0; 270 271 audit_enabled = (*audit_vpp != NULL); 272 273 /* 274 * XXX: What to do about write failures here? 275 */ 276 if (old_vp != NULL) { 277 AUDIT_PRINTF(("Closing old audit file\n")); 278 mtx_unlock(&audit_mtx); 279 vfslocked = VFS_LOCK_GIANT(old_vp->v_mount); 280 vn_close(old_vp, AUDIT_CLOSE_FLAGS, old_cred, 281 audit_td); 282 VFS_UNLOCK_GIANT(vfslocked); 283 crfree(old_cred); 284 mtx_lock(&audit_mtx); 285 old_cred = NULL; 286 old_vp = NULL; 287 AUDIT_PRINTF(("Audit file closed\n")); 288 } 289 if (*audit_vpp != NULL) { 290 AUDIT_PRINTF(("Opening new audit file\n")); 291 } 292 do_replacement_signal = 1; 293 } 294 295 /* 296 * Signal that replacement have occurred to wake up and 297 * start any other replacements started in parallel. We can 298 * continue about our business in the mean time. We 299 * broadcast so that both new replacements can be inserted, 300 * but also so that the source(s) of replacement can return 301 * successfully. 302 */ 303 if (do_replacement_signal) 304 cv_broadcast(&audit_replacement_cv); 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 */ 314 static void 315 audit_worker_process_record(struct vnode *audit_vp, struct ucred *audit_cred, 316 struct thread *audit_td, struct kaudit_record *ar) 317 { 318 struct au_record *bsm; 319 au_class_t class; 320 au_event_t event; 321 int error, ret; 322 au_id_t auid; 323 int sorf; 324 325 if ((ar->k_ar_commit & AR_COMMIT_USER) && 326 (ar->k_ar_commit & AR_PRESELECT_TRAIL)) { 327 error = audit_record_write(audit_vp, audit_cred, audit_td, 328 ar->k_udata, ar->k_ulen); 329 if (error && audit_panic_on_write_fail) 330 panic("audit_worker: write error %d\n", error); 331 else if (error) 332 printf("audit_worker: write error %d\n", error); 333 } 334 if ((ar->k_ar_commit & AR_COMMIT_USER) && 335 (ar->k_ar_commit & AR_PRESELECT_PIPE)) 336 audit_pipe_submit_user(ar->k_udata, ar->k_ulen); 337 338 if (!(ar->k_ar_commit & AR_COMMIT_KERNEL)) 339 return; 340 341 auid = ar->k_ar.ar_subj_auid; 342 event = ar->k_ar.ar_event; 343 class = au_event_class(event); 344 if (ar->k_ar.ar_errno == 0) 345 sorf = AU_PRS_SUCCESS; 346 else 347 sorf = AU_PRS_FAILURE; 348 349 ret = kaudit_to_bsm(ar, &bsm); 350 switch (ret) { 351 case BSM_NOAUDIT: 352 return; 353 354 case BSM_FAILURE: 355 printf("audit_worker_process_record: BSM_FAILURE\n"); 356 return; 357 358 case BSM_SUCCESS: 359 break; 360 361 default: 362 panic("kaudit_to_bsm returned %d", ret); 363 } 364 365 if (ar->k_ar_commit & AR_PRESELECT_TRAIL) { 366 error = audit_record_write(audit_vp, audit_cred, 367 audit_td, bsm->data, bsm->len); 368 if (error && audit_panic_on_write_fail) 369 panic("audit_worker: write error %d\n", 370 error); 371 else if (error) 372 printf("audit_worker: write error %d\n", 373 error); 374 } 375 if (ar->k_ar_commit & AR_PRESELECT_PIPE) 376 audit_pipe_submit(auid, event, class, sorf, 377 ar->k_ar_commit & AR_PRESELECT_TRAIL, bsm->data, 378 bsm->len); 379 kau_free(bsm); 380 } 381 382 /* 383 * The audit_worker thread is responsible for watching the event queue, 384 * dequeueing records, converting them to BSM format, and committing them to 385 * disk. In order to minimize lock thrashing, records are dequeued in sets 386 * to a thread-local work queue. In addition, the audit_work performs the 387 * actual exchange of audit log vnode pointer, as audit_vp is a thread-local 388 * variable. 389 */ 390 static void 391 audit_worker(void *arg) 392 { 393 struct kaudit_queue ar_worklist; 394 struct kaudit_record *ar; 395 struct ucred *audit_cred; 396 struct thread *audit_td; 397 struct vnode *audit_vp; 398 int lowater_signal; 399 400 AUDIT_PRINTF(("audit_worker starting\n")); 401 402 /* 403 * These are thread-local variables requiring no synchronization. 404 */ 405 TAILQ_INIT(&ar_worklist); 406 audit_cred = NULL; 407 audit_td = curthread; 408 audit_vp = NULL; 409 410 mtx_lock(&audit_mtx); 411 while (1) { 412 mtx_assert(&audit_mtx, MA_OWNED); 413 414 /* 415 * Wait for record or rotation events. 416 */ 417 while (!audit_replacement_flag && TAILQ_EMPTY(&audit_q)) { 418 AUDIT_PRINTF(("audit_worker waiting\n")); 419 cv_wait(&audit_worker_cv, &audit_mtx); 420 AUDIT_PRINTF(("audit_worker woken up\n")); 421 AUDIT_PRINTF(("audit_worker: new vp = %p; value of " 422 "flag %d\n", audit_replacement_vp, 423 audit_replacement_flag)); 424 } 425 426 /* 427 * First priority: replace the audit log target if requested. 428 */ 429 audit_worker_rotate(&audit_cred, &audit_vp, audit_td); 430 431 /* 432 * If there are records in the global audit record queue, 433 * transfer them to a thread-local queue and process them 434 * one by one. If we cross the low watermark threshold, 435 * signal any waiting processes that they may wake up and 436 * continue generating records. 437 */ 438 lowater_signal = 0; 439 while ((ar = TAILQ_FIRST(&audit_q))) { 440 TAILQ_REMOVE(&audit_q, ar, k_q); 441 audit_q_len--; 442 if (audit_q_len == audit_qctrl.aq_lowater) 443 lowater_signal++; 444 TAILQ_INSERT_TAIL(&ar_worklist, ar, k_q); 445 } 446 if (lowater_signal) 447 cv_broadcast(&audit_watermark_cv); 448 449 mtx_unlock(&audit_mtx); 450 while ((ar = TAILQ_FIRST(&ar_worklist))) { 451 TAILQ_REMOVE(&ar_worklist, ar, k_q); 452 audit_worker_process_record(audit_vp, audit_cred, 453 audit_td, ar); 454 audit_free(ar); 455 } 456 mtx_lock(&audit_mtx); 457 } 458 } 459 460 /* 461 * audit_rotate_vnode() is called by a user or kernel thread to configure or 462 * de-configure auditing on a vnode. The arguments are the replacement 463 * credential and vnode to substitute for the current credential and vnode, 464 * if any. If either is set to NULL, both should be NULL, and this is used 465 * to indicate that audit is being disabled. The real work is done in the 466 * audit_worker thread, but audit_rotate_vnode() waits synchronously for that 467 * to complete. 468 * 469 * The vnode should be referenced and opened by the caller. The credential 470 * should be referenced. audit_rotate_vnode() will own both references as of 471 * this call, so the caller should not release either. 472 * 473 * XXXAUDIT: Review synchronize communication logic. Really, this is a 474 * message queue of depth 1. 475 * 476 * XXXAUDIT: Enhance the comments below to indicate that we are basically 477 * acquiring ownership of the communications queue, inserting our message, 478 * and waiting for an acknowledgement. 479 */ 480 void 481 audit_rotate_vnode(struct ucred *cred, struct vnode *vp) 482 { 483 484 /* 485 * If other parallel log replacements have been requested, we wait 486 * until they've finished before continuing. 487 */ 488 mtx_lock(&audit_mtx); 489 while (audit_replacement_flag != 0) { 490 AUDIT_PRINTF(("audit_rotate_vnode: sleeping to wait for " 491 "flag\n")); 492 cv_wait(&audit_replacement_cv, &audit_mtx); 493 AUDIT_PRINTF(("audit_rotate_vnode: woken up (flag %d)\n", 494 audit_replacement_flag)); 495 } 496 audit_replacement_cred = cred; 497 audit_replacement_flag = 1; 498 audit_replacement_vp = vp; 499 500 /* 501 * Wake up the audit worker to perform the exchange once we 502 * release the mutex. 503 */ 504 cv_signal(&audit_worker_cv); 505 506 /* 507 * Wait for the audit_worker to broadcast that a replacement has 508 * taken place; we know that once this has happened, our vnode 509 * has been replaced in, so we can return successfully. 510 */ 511 AUDIT_PRINTF(("audit_rotate_vnode: waiting for news of " 512 "replacement\n")); 513 cv_wait(&audit_replacement_cv, &audit_mtx); 514 AUDIT_PRINTF(("audit_rotate_vnode: change acknowledged by " 515 "audit_worker (flag " "now %d)\n", audit_replacement_flag)); 516 mtx_unlock(&audit_mtx); 517 518 audit_file_rotate_wait = 0; /* We can now request another rotation */ 519 } 520 521 void 522 audit_worker_init(void) 523 { 524 int error; 525 526 cv_init(&audit_replacement_cv, "audit_replacement_cv"); 527 error = kthread_create(audit_worker, NULL, &audit_thread, RFHIGHPID, 528 0, "audit_worker"); 529 if (error) 530 panic("audit_worker_init: kthread_create returned %d", error); 531 } 532