audit_worker.c revision 159263
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: head/sys/security/audit/audit_worker.c 159263 2006-06-05 13:46:55Z rwatson $ 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 */ 75static 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 */ 93static struct cv audit_replacement_cv; 94 95static int audit_replacement_flag; 96static struct vnode *audit_replacement_vp; 97static struct ucred *audit_replacement_cred; 98 99/* 100 * Flags related to Kernel->user-space communication. 101 */ 102static 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 */ 111static int 112audit_record_write(struct vnode *vp, struct kaudit_record *ar, 113 struct ucred *cred, struct thread *td) 114{ 115 int ret; 116 long temp; 117 struct au_record *bsm; 118 struct vattr vattr; 119 struct statfs *mnt_stat = &vp->v_mount->mnt_stat; 120 int vfslocked; 121 122 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 123 124 /* 125 * First, gather statistics on the audit log file and file system so 126 * that we know how we're doing on space. In both cases, if we're 127 * unable to perform the operation, we drop the record and return. 128 * However, this is arguably an assertion failure. 129 * XXX Need a FreeBSD equivalent. 130 */ 131 ret = VFS_STATFS(vp->v_mount, mnt_stat, td); 132 if (ret) 133 goto out; 134 135 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td); 136 ret = VOP_GETATTR(vp, &vattr, cred, td); 137 VOP_UNLOCK(vp, 0, td); 138 if (ret) 139 goto out; 140 141 /* update the global stats struct */ 142 audit_fstat.af_currsz = vattr.va_size; 143 144 /* 145 * XXX Need to decide what to do if the trigger to the audit daemon 146 * fails. 147 */ 148 149 /* 150 * If we fall below minimum free blocks (hard limit), tell the audit 151 * daemon to force a rotation off of the file system. We also stop 152 * writing, which means this audit record is probably lost. If we 153 * fall below the minimum percent free blocks (soft limit), then 154 * kindly suggest to the audit daemon to do something. 155 */ 156 if (mnt_stat->f_bfree < AUDIT_HARD_LIMIT_FREE_BLOCKS) { 157 (void)send_trigger(AUDIT_TRIGGER_NO_SPACE); 158 /* 159 * Hopefully userspace did something about all the previous 160 * triggers that were sent prior to this critical condition. 161 * If fail-stop is set, then we're done; goodnight Gracie. 162 */ 163 if (audit_fail_stop) 164 panic("Audit log space exhausted and fail-stop set."); 165 else { 166 audit_suspended = 1; 167 ret = ENOSPC; 168 goto out; 169 } 170 } else 171 /* 172 * Send a message to the audit daemon that disk space is 173 * getting low. 174 * 175 * XXXAUDIT: Check math and block size calculation here. 176 */ 177 if (audit_qctrl.aq_minfree != 0) { 178 temp = mnt_stat->f_blocks / (100 / 179 audit_qctrl.aq_minfree); 180 if (mnt_stat->f_bfree < temp) 181 (void)send_trigger(AUDIT_TRIGGER_LOW_SPACE); 182 } 183 184 /* 185 * Check if the current log file is full; if so, call for a log 186 * rotate. This is not an exact comparison; we may write some records 187 * over the limit. If that's not acceptable, then add a fudge factor 188 * here. 189 */ 190 if ((audit_fstat.af_filesz != 0) && 191 (audit_file_rotate_wait == 0) && 192 (vattr.va_size >= audit_fstat.af_filesz)) { 193 audit_file_rotate_wait = 1; 194 (void)send_trigger(AUDIT_TRIGGER_OPEN_NEW); 195 } 196 197 /* 198 * If the estimated amount of audit data in the audit event queue 199 * (plus records allocated but not yet queued) has reached the amount 200 * of free space on the disk, then we need to go into an audit fail 201 * stop state, in which we do not permit the allocation/committing of 202 * any new audit records. We continue to process packets but don't 203 * allow any activities that might generate new records. In the 204 * future, we might want to detect when space is available again and 205 * allow operation to continue, but this behavior is sufficient to 206 * meet fail stop requirements in CAPP. 207 */ 208 if (audit_fail_stop && 209 (unsigned long) 210 ((audit_q_len + audit_pre_q_len + 1) * MAX_AUDIT_RECORD_SIZE) / 211 mnt_stat->f_bsize >= (unsigned long)(mnt_stat->f_bfree)) { 212 printf("audit_record_write: free space below size of audit " 213 "queue, failing stop\n"); 214 audit_in_failure = 1; 215 } 216 217 /* 218 * If there is a user audit record attached to the kernel record, 219 * then write the user record. 220 * 221 * XXX Need to decide a few things here: IF the user audit record is 222 * written, but the write of the kernel record fails, what to do? 223 * Should the kernel record come before or after the user record? 224 * For now, we write the user record first, and we ignore errors. 225 */ 226 if (ar->k_ar_commit & AR_COMMIT_USER) { 227 /* 228 * Try submitting the record to any active audit pipes. 229 */ 230 audit_pipe_submit((void *)ar->k_udata, ar->k_ulen); 231 232 /* 233 * And to disk. 234 */ 235 ret = vn_rdwr(UIO_WRITE, vp, (void *)ar->k_udata, ar->k_ulen, 236 (off_t)0, UIO_SYSSPACE, IO_APPEND|IO_UNIT, cred, NULL, 237 NULL, td); 238 if (ret) 239 goto out; 240 } 241 242 /* 243 * Convert the internal kernel record to BSM format and write it out 244 * if everything's OK. 245 */ 246 if (!(ar->k_ar_commit & AR_COMMIT_KERNEL)) { 247 ret = 0; 248 goto out; 249 } 250 251 /* 252 * XXXAUDIT: Should we actually allow this conversion to fail? With 253 * sleeping memory allocation and invariants checks, perhaps not. 254 */ 255 ret = kaudit_to_bsm(ar, &bsm); 256 if (ret == BSM_NOAUDIT) { 257 ret = 0; 258 goto out; 259 } 260 261 /* 262 * XXX: We drop the record on BSM conversion failure, but really this 263 * is an assertion failure. 264 */ 265 if (ret == BSM_FAILURE) { 266 AUDIT_PRINTF(("BSM conversion failure\n")); 267 ret = EINVAL; 268 goto out; 269 } 270 271 /* 272 * Try submitting the record to any active audit pipes. 273 */ 274 audit_pipe_submit((void *)bsm->data, bsm->len); 275 276 /* 277 * XXX We should break the write functionality away from the BSM 278 * record generation and have the BSM generation done before this 279 * function is called. This function will then take the BSM record as 280 * a parameter. 281 */ 282 ret = (vn_rdwr(UIO_WRITE, vp, (void *)bsm->data, bsm->len, (off_t)0, 283 UIO_SYSSPACE, IO_APPEND|IO_UNIT, cred, NULL, NULL, td)); 284 kau_free(bsm); 285 286out: 287 /* 288 * When we're done processing the current record, we have to check to 289 * see if we're in a failure mode, and if so, whether this was the 290 * last record left to be drained. If we're done draining, then we 291 * fsync the vnode and panic. 292 */ 293 if (audit_in_failure && audit_q_len == 0 && audit_pre_q_len == 0) { 294 VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td); 295 (void)VOP_FSYNC(vp, MNT_WAIT, td); 296 VOP_UNLOCK(vp, 0, td); 297 panic("Audit store overflow; record queue drained."); 298 } 299 300 VFS_UNLOCK_GIANT(vfslocked); 301 302 return (ret); 303} 304 305/* 306 * If an appropriate signal has been received rotate the audit log based on 307 * the global replacement variables. Signal consumers as needed that the 308 * rotation has taken place. 309 * 310 * XXXRW: The global variables and CVs used to signal the audit_worker to 311 * perform a rotation are essentially a message queue of depth 1. It would 312 * be much nicer to actually use a message queue. 313 */ 314static void 315audit_worker_rotate(struct ucred **audit_credp, struct vnode **audit_vpp, 316 struct thread *audit_td) 317{ 318 int do_replacement_signal, vfslocked; 319 struct ucred *old_cred; 320 struct vnode *old_vp; 321 322 mtx_assert(&audit_mtx, MA_OWNED); 323 324 do_replacement_signal = 0; 325 while (audit_replacement_flag != 0) { 326 old_cred = *audit_credp; 327 old_vp = *audit_vpp; 328 *audit_credp = audit_replacement_cred; 329 *audit_vpp = audit_replacement_vp; 330 audit_replacement_cred = NULL; 331 audit_replacement_vp = NULL; 332 audit_replacement_flag = 0; 333 334 audit_enabled = (*audit_vpp != NULL); 335 336 /* 337 * XXX: What to do about write failures here? 338 */ 339 if (old_vp != NULL) { 340 AUDIT_PRINTF(("Closing old audit file\n")); 341 mtx_unlock(&audit_mtx); 342 vfslocked = VFS_LOCK_GIANT(old_vp->v_mount); 343 vn_close(old_vp, AUDIT_CLOSE_FLAGS, old_cred, 344 audit_td); 345 VFS_UNLOCK_GIANT(vfslocked); 346 crfree(old_cred); 347 mtx_lock(&audit_mtx); 348 old_cred = NULL; 349 old_vp = NULL; 350 AUDIT_PRINTF(("Audit file closed\n")); 351 } 352 if (*audit_vpp != NULL) { 353 AUDIT_PRINTF(("Opening new audit file\n")); 354 } 355 do_replacement_signal = 1; 356 } 357 358 /* 359 * Signal that replacement have occurred to wake up and 360 * start any other replacements started in parallel. We can 361 * continue about our business in the mean time. We 362 * broadcast so that both new replacements can be inserted, 363 * but also so that the source(s) of replacement can return 364 * successfully. 365 */ 366 if (do_replacement_signal) 367 cv_broadcast(&audit_replacement_cv); 368} 369 370/* 371 * Drain the audit commit queue and free the records. Used if there are 372 * records present, but no audit log target. 373 */ 374static void 375audit_worker_drain(void) 376{ 377 struct kaudit_record *ar; 378 379 mtx_assert(&audit_mtx, MA_OWNED); 380 381 while ((ar = TAILQ_FIRST(&audit_q))) { 382 TAILQ_REMOVE(&audit_q, ar, k_q); 383 audit_free(ar); 384 audit_q_len--; 385 } 386} 387 388/* 389 * Given a kernel audit record, process as required. Currently, that means 390 * passing it to audit_record_write(), but in the future it will mean 391 * converting it to BSM and then routing it to various possible output 392 * streams, including the audit trail and audit pipes. The caller will free 393 * the record. 394 */ 395static void 396audit_worker_process_record(struct vnode *audit_vp, struct ucred *audit_cred, 397 struct thread *audit_td, struct kaudit_record *ar) 398{ 399 int error; 400 401 if (audit_vp == NULL) 402 return; 403 404 error = audit_record_write(audit_vp, ar, audit_cred, audit_td); 405 if (error) { 406 if (audit_panic_on_write_fail) 407 panic("audit_worker: write error %d\n", error); 408 else 409 printf("audit_worker: write error %d\n", error); 410 } 411} 412 413/* 414 * The audit_worker thread is responsible for watching the event queue, 415 * dequeueing records, converting them to BSM format, and committing them to 416 * disk. In order to minimize lock thrashing, records are dequeued in sets 417 * to a thread-local work queue. In addition, the audit_work performs the 418 * actual exchange of audit log vnode pointer, as audit_vp is a thread-local 419 * variable. 420 */ 421static void 422audit_worker(void *arg) 423{ 424 struct kaudit_queue ar_worklist; 425 struct kaudit_record *ar; 426 struct ucred *audit_cred; 427 struct thread *audit_td; 428 struct vnode *audit_vp; 429 int lowater_signal; 430 431 AUDIT_PRINTF(("audit_worker starting\n")); 432 433 /* 434 * These are thread-local variables requiring no synchronization. 435 */ 436 TAILQ_INIT(&ar_worklist); 437 audit_cred = NULL; 438 audit_td = curthread; 439 audit_vp = NULL; 440 441 mtx_lock(&audit_mtx); 442 while (1) { 443 mtx_assert(&audit_mtx, MA_OWNED); 444 445 /* 446 * Wait for record or rotation events. 447 */ 448 while (!audit_replacement_flag && TAILQ_EMPTY(&audit_q)) { 449 AUDIT_PRINTF(("audit_worker waiting\n")); 450 cv_wait(&audit_worker_cv, &audit_mtx); 451 AUDIT_PRINTF(("audit_worker woken up\n")); 452 AUDIT_PRINTF(("audit_worker: new vp = %p; value of " 453 "flag %d\n", audit_replacement_vp, 454 audit_replacement_flag)); 455 } 456 457 /* 458 * First priority: replace the audit log target if requested. 459 */ 460 audit_worker_rotate(&audit_cred, &audit_vp, audit_td); 461 462 /* 463 * If we have records, but there's no active vnode to write 464 * to, drain the record queue. Generally, we prevent the 465 * unnecessary allocation of records elsewhere, but we need 466 * to allow for races between conditional allocation and 467 * queueing. Go back to waiting when we're done. 468 */ 469 if (audit_vp == NULL) { 470 audit_worker_drain(); 471 continue; 472 } 473 474 /* 475 * We have both records to write and an active vnode to write 476 * to. Dequeue a record, and start the write. Eventually, 477 * it might make sense to dequeue several records and perform 478 * our own clustering, if the lower layers aren't doing it 479 * automatically enough. 480 */ 481 lowater_signal = 0; 482 while ((ar = TAILQ_FIRST(&audit_q))) { 483 TAILQ_REMOVE(&audit_q, ar, k_q); 484 audit_q_len--; 485 if (audit_q_len == audit_qctrl.aq_lowater) 486 lowater_signal++; 487 TAILQ_INSERT_TAIL(&ar_worklist, ar, k_q); 488 } 489 if (lowater_signal) 490 cv_broadcast(&audit_watermark_cv); 491 492 mtx_unlock(&audit_mtx); 493 while ((ar = TAILQ_FIRST(&ar_worklist))) { 494 TAILQ_REMOVE(&ar_worklist, ar, k_q); 495 audit_worker_process_record(audit_vp, audit_cred, 496 audit_td, ar); 497 audit_free(ar); 498 } 499 mtx_lock(&audit_mtx); 500 } 501} 502 503/* 504 * audit_rotate_vnode() is called by a user or kernel thread to configure or 505 * de-configure auditing on a vnode. The arguments are the replacement 506 * credential and vnode to substitute for the current credential and vnode, 507 * if any. If either is set to NULL, both should be NULL, and this is used 508 * to indicate that audit is being disabled. The real work is done in the 509 * audit_worker thread, but audit_rotate_vnode() waits synchronously for that 510 * to complete. 511 * 512 * The vnode should be referenced and opened by the caller. The credential 513 * should be referenced. audit_rotate_vnode() will own both references as of 514 * this call, so the caller should not release either. 515 * 516 * XXXAUDIT: Review synchronize communication logic. Really, this is a 517 * message queue of depth 1. 518 * 519 * XXXAUDIT: Enhance the comments below to indicate that we are basically 520 * acquiring ownership of the communications queue, inserting our message, 521 * and waiting for an acknowledgement. 522 */ 523void 524audit_rotate_vnode(struct ucred *cred, struct vnode *vp) 525{ 526 527 /* 528 * If other parallel log replacements have been requested, we wait 529 * until they've finished before continuing. 530 */ 531 mtx_lock(&audit_mtx); 532 while (audit_replacement_flag != 0) { 533 AUDIT_PRINTF(("audit_rotate_vnode: sleeping to wait for " 534 "flag\n")); 535 cv_wait(&audit_replacement_cv, &audit_mtx); 536 AUDIT_PRINTF(("audit_rotate_vnode: woken up (flag %d)\n", 537 audit_replacement_flag)); 538 } 539 audit_replacement_cred = cred; 540 audit_replacement_flag = 1; 541 audit_replacement_vp = vp; 542 543 /* 544 * Wake up the audit worker to perform the exchange once we 545 * release the mutex. 546 */ 547 cv_signal(&audit_worker_cv); 548 549 /* 550 * Wait for the audit_worker to broadcast that a replacement has 551 * taken place; we know that once this has happened, our vnode 552 * has been replaced in, so we can return successfully. 553 */ 554 AUDIT_PRINTF(("audit_rotate_vnode: waiting for news of " 555 "replacement\n")); 556 cv_wait(&audit_replacement_cv, &audit_mtx); 557 AUDIT_PRINTF(("audit_rotate_vnode: change acknowledged by " 558 "audit_worker (flag " "now %d)\n", audit_replacement_flag)); 559 mtx_unlock(&audit_mtx); 560 561 audit_file_rotate_wait = 0; /* We can now request another rotation */ 562} 563 564void 565audit_worker_init(void) 566{ 567 int error; 568 569 cv_init(&audit_replacement_cv, "audit_replacement_cv"); 570 error = kthread_create(audit_worker, NULL, &audit_thread, RFHIGHPID, 571 0, "audit_worker"); 572 if (error) 573 panic("audit_worker_init: kthread_create returned %d", error); 574} 575