1 /* 2 * linux/kernel/acct.c 3 * 4 * BSD Process Accounting for Linux 5 * 6 * Author: Marco van Wieringen <mvw@planets.elm.net> 7 * 8 * Some code based on ideas and code from: 9 * Thomas K. Dyas <tdyas@eden.rutgers.edu> 10 * 11 * This file implements BSD-style process accounting. Whenever any 12 * process exits, an accounting record of type "struct acct" is 13 * written to the file specified with the acct() system call. It is 14 * up to user-level programs to do useful things with the accounting 15 * log. The kernel just provides the raw accounting information. 16 * 17 * (C) Copyright 1995 - 1997 Marco van Wieringen - ELM Consultancy B.V. 18 * 19 * Plugged two leaks. 1) It didn't return acct_file into the free_filps if 20 * the file happened to be read-only. 2) If the accounting was suspended 21 * due to the lack of space it happily allowed to reopen it and completely 22 * lost the old acct_file. 3/10/98, Al Viro. 23 * 24 * Now we silently close acct_file on attempt to reopen. Cleaned sys_acct(). 25 * XTerms and EMACS are manifestations of pure evil. 21/10/98, AV. 26 * 27 * Fixed a nasty interaction with with sys_umount(). If the accointing 28 * was suspeneded we failed to stop it on umount(). Messy. 29 * Another one: remount to readonly didn't stop accounting. 30 * Question: what should we do if we have CAP_SYS_ADMIN but not 31 * CAP_SYS_PACCT? Current code does the following: umount returns -EBUSY 32 * unless we are messing with the root. In that case we are getting a 33 * real mess with do_remount_sb(). 9/11/98, AV. 34 * 35 * Fixed a bunch of races (and pair of leaks). Probably not the best way, 36 * but this one obviously doesn't introduce deadlocks. Later. BTW, found 37 * one race (and leak) in BSD implementation. 38 * OK, that's better. ANOTHER race and leak in BSD variant. There always 39 * is one more bug... 10/11/98, AV. 40 * 41 * Oh, fsck... Oopsable SMP race in do_process_acct() - we must hold 42 * ->mmap_sem to walk the vma list of current->mm. Nasty, since it leaks 43 * a struct file opened for write. Fixed. 2/6/2000, AV. 44 */ 45 46 #include <linux/mm.h> 47 #include <linux/slab.h> 48 #include <linux/acct.h> 49 #include <linux/capability.h> 50 #include <linux/file.h> 51 #include <linux/tty.h> 52 #include <linux/security.h> 53 #include <linux/vfs.h> 54 #include <linux/jiffies.h> 55 #include <linux/times.h> 56 #include <linux/syscalls.h> 57 #include <linux/mount.h> 58 #include <linux/uaccess.h> 59 #include <asm/div64.h> 60 #include <linux/blkdev.h> /* sector_div */ 61 #include <linux/pid_namespace.h> 62 #include <../fs/mount.h> /* will go away when we refactor */ 63 64 /* 65 * These constants control the amount of freespace that suspend and 66 * resume the process accounting system, and the time delay between 67 * each check. 68 * Turned into sysctl-controllable parameters. AV, 12/11/98 69 */ 70 71 int acct_parm[3] = {4, 2, 30}; 72 #define RESUME (acct_parm[0]) /* >foo% free space - resume */ 73 #define SUSPEND (acct_parm[1]) /* <foo% free space - suspend */ 74 #define ACCT_TIMEOUT (acct_parm[2]) /* foo second timeout between checks */ 75 76 /* 77 * External references and all of the globals. 78 */ 79 static void do_acct_process(struct bsd_acct_struct *acct); 80 81 struct bsd_acct_struct { 82 atomic_long_t count; 83 union { 84 struct { 85 struct hlist_node s_list; 86 struct hlist_node m_list; 87 }; 88 struct rcu_head rcu; 89 }; 90 struct mutex lock; 91 int active; 92 unsigned long needcheck; 93 struct file *file; 94 struct pid_namespace *ns; 95 struct work_struct work; 96 struct completion done; 97 }; 98 99 static void acct_free_rcu(struct rcu_head *head) 100 { 101 kfree(container_of(head, struct bsd_acct_struct, rcu)); 102 } 103 104 static DEFINE_SPINLOCK(acct_lock); 105 106 /* 107 * Check the amount of free space and suspend/resume accordingly. 108 */ 109 static int check_free_space(struct bsd_acct_struct *acct) 110 { 111 struct kstatfs sbuf; 112 113 if (time_is_before_jiffies(acct->needcheck)) 114 goto out; 115 116 /* May block */ 117 if (vfs_statfs(&acct->file->f_path, &sbuf)) 118 goto out; 119 120 if (acct->active) { 121 u64 suspend = sbuf.f_blocks * SUSPEND; 122 do_div(suspend, 100); 123 if (sbuf.f_bavail <= suspend) { 124 acct->active = 0; 125 printk(KERN_INFO "Process accounting paused\n"); 126 } 127 } else { 128 u64 resume = sbuf.f_blocks * RESUME; 129 do_div(resume, 100); 130 if (sbuf.f_bavail >= resume) { 131 acct->active = 1; 132 printk(KERN_INFO "Process accounting resumed\n"); 133 } 134 } 135 136 acct->needcheck = jiffies + ACCT_TIMEOUT*HZ; 137 out: 138 return acct->active; 139 } 140 141 static void acct_put(struct bsd_acct_struct *p) 142 { 143 if (atomic_long_dec_and_test(&p->count)) 144 call_rcu(&p->rcu, acct_free_rcu); 145 } 146 147 static struct bsd_acct_struct *__acct_get(struct bsd_acct_struct *res) 148 { 149 if (!atomic_long_inc_not_zero(&res->count)) { 150 rcu_read_unlock(); 151 cpu_relax(); 152 return NULL; 153 } 154 rcu_read_unlock(); 155 mutex_lock(&res->lock); 156 if (!res->ns) { 157 mutex_unlock(&res->lock); 158 acct_put(res); 159 return NULL; 160 } 161 return res; 162 } 163 164 static struct bsd_acct_struct *acct_get(struct pid_namespace *ns) 165 { 166 struct bsd_acct_struct *res; 167 again: 168 smp_rmb(); 169 rcu_read_lock(); 170 res = ACCESS_ONCE(ns->bacct); 171 if (!res) { 172 rcu_read_unlock(); 173 return NULL; 174 } 175 res = __acct_get(res); 176 if (!res) 177 goto again; 178 return res; 179 } 180 181 static void close_work(struct work_struct *work) 182 { 183 struct bsd_acct_struct *acct = container_of(work, struct bsd_acct_struct, work); 184 struct file *file = acct->file; 185 mnt_unpin(file->f_path.mnt); 186 if (file->f_op->flush) 187 file->f_op->flush(file, NULL); 188 __fput_sync(file); 189 complete(&acct->done); 190 } 191 192 static void acct_kill(struct bsd_acct_struct *acct, 193 struct bsd_acct_struct *new) 194 { 195 if (acct) { 196 struct pid_namespace *ns = acct->ns; 197 do_acct_process(acct); 198 INIT_WORK(&acct->work, close_work); 199 init_completion(&acct->done); 200 schedule_work(&acct->work); 201 wait_for_completion(&acct->done); 202 spin_lock(&acct_lock); 203 hlist_del(&acct->m_list); 204 hlist_del(&acct->s_list); 205 spin_unlock(&acct_lock); 206 ns->bacct = new; 207 if (new) { 208 struct vfsmount *m = new->file->f_path.mnt; 209 mnt_pin(m); 210 spin_lock(&acct_lock); 211 hlist_add_head(&new->s_list, &m->mnt_sb->s_pins); 212 hlist_add_head(&new->m_list, &real_mount(m)->mnt_pins); 213 spin_unlock(&acct_lock); 214 mutex_unlock(&new->lock); 215 } 216 acct->ns = NULL; 217 atomic_long_dec(&acct->count); 218 mutex_unlock(&acct->lock); 219 acct_put(acct); 220 } 221 } 222 223 static int acct_on(struct filename *pathname) 224 { 225 struct file *file; 226 struct vfsmount *mnt; 227 struct pid_namespace *ns = task_active_pid_ns(current); 228 struct bsd_acct_struct *acct, *old; 229 230 acct = kzalloc(sizeof(struct bsd_acct_struct), GFP_KERNEL); 231 if (!acct) 232 return -ENOMEM; 233 234 /* Difference from BSD - they don't do O_APPEND */ 235 file = file_open_name(pathname, O_WRONLY|O_APPEND|O_LARGEFILE, 0); 236 if (IS_ERR(file)) { 237 kfree(acct); 238 return PTR_ERR(file); 239 } 240 241 if (!S_ISREG(file_inode(file)->i_mode)) { 242 kfree(acct); 243 filp_close(file, NULL); 244 return -EACCES; 245 } 246 247 if (!file->f_op->write) { 248 kfree(acct); 249 filp_close(file, NULL); 250 return -EIO; 251 } 252 253 atomic_long_set(&acct->count, 1); 254 acct->file = file; 255 acct->needcheck = jiffies; 256 acct->ns = ns; 257 mutex_init(&acct->lock); 258 mnt = file->f_path.mnt; 259 260 old = acct_get(ns); 261 mutex_lock_nested(&acct->lock, 1); /* nobody has seen it yet */ 262 if (old) { 263 acct_kill(old, acct); 264 } else { 265 ns->bacct = acct; 266 spin_lock(&acct_lock); 267 mnt_pin(mnt); 268 hlist_add_head(&acct->s_list, &mnt->mnt_sb->s_pins); 269 hlist_add_head(&acct->m_list, &real_mount(mnt)->mnt_pins); 270 spin_unlock(&acct_lock); 271 mutex_unlock(&acct->lock); 272 } 273 mntput(mnt); /* it's pinned, now give up active reference */ 274 return 0; 275 } 276 277 static DEFINE_MUTEX(acct_on_mutex); 278 279 /** 280 * sys_acct - enable/disable process accounting 281 * @name: file name for accounting records or NULL to shutdown accounting 282 * 283 * Returns 0 for success or negative errno values for failure. 284 * 285 * sys_acct() is the only system call needed to implement process 286 * accounting. It takes the name of the file where accounting records 287 * should be written. If the filename is NULL, accounting will be 288 * shutdown. 289 */ 290 SYSCALL_DEFINE1(acct, const char __user *, name) 291 { 292 int error = 0; 293 294 if (!capable(CAP_SYS_PACCT)) 295 return -EPERM; 296 297 if (name) { 298 struct filename *tmp = getname(name); 299 if (IS_ERR(tmp)) 300 return PTR_ERR(tmp); 301 mutex_lock(&acct_on_mutex); 302 error = acct_on(tmp); 303 mutex_unlock(&acct_on_mutex); 304 putname(tmp); 305 } else { 306 acct_kill(acct_get(task_active_pid_ns(current)), NULL); 307 } 308 309 return error; 310 } 311 312 void acct_auto_close_mnt(struct hlist_head *list) 313 { 314 rcu_read_lock(); 315 while (1) { 316 struct hlist_node *p = ACCESS_ONCE(list->first); 317 if (!p) 318 break; 319 acct_kill(__acct_get(hlist_entry(p, 320 struct bsd_acct_struct, 321 m_list)), NULL); 322 rcu_read_lock(); 323 } 324 rcu_read_unlock(); 325 } 326 327 void acct_auto_close(struct hlist_head *list) 328 { 329 rcu_read_lock(); 330 while (1) { 331 struct hlist_node *p = ACCESS_ONCE(list->first); 332 if (!p) 333 break; 334 acct_kill(__acct_get(hlist_entry(p, 335 struct bsd_acct_struct, 336 s_list)), NULL); 337 rcu_read_lock(); 338 } 339 rcu_read_unlock(); 340 } 341 342 void acct_exit_ns(struct pid_namespace *ns) 343 { 344 acct_kill(acct_get(ns), NULL); 345 } 346 347 /* 348 * encode an unsigned long into a comp_t 349 * 350 * This routine has been adopted from the encode_comp_t() function in 351 * the kern_acct.c file of the FreeBSD operating system. The encoding 352 * is a 13-bit fraction with a 3-bit (base 8) exponent. 353 */ 354 355 #define MANTSIZE 13 /* 13 bit mantissa. */ 356 #define EXPSIZE 3 /* Base 8 (3 bit) exponent. */ 357 #define MAXFRACT ((1 << MANTSIZE) - 1) /* Maximum fractional value. */ 358 359 static comp_t encode_comp_t(unsigned long value) 360 { 361 int exp, rnd; 362 363 exp = rnd = 0; 364 while (value > MAXFRACT) { 365 rnd = value & (1 << (EXPSIZE - 1)); /* Round up? */ 366 value >>= EXPSIZE; /* Base 8 exponent == 3 bit shift. */ 367 exp++; 368 } 369 370 /* 371 * If we need to round up, do it (and handle overflow correctly). 372 */ 373 if (rnd && (++value > MAXFRACT)) { 374 value >>= EXPSIZE; 375 exp++; 376 } 377 378 /* 379 * Clean it up and polish it off. 380 */ 381 exp <<= MANTSIZE; /* Shift the exponent into place */ 382 exp += value; /* and add on the mantissa. */ 383 return exp; 384 } 385 386 #if ACCT_VERSION==1 || ACCT_VERSION==2 387 /* 388 * encode an u64 into a comp2_t (24 bits) 389 * 390 * Format: 5 bit base 2 exponent, 20 bits mantissa. 391 * The leading bit of the mantissa is not stored, but implied for 392 * non-zero exponents. 393 * Largest encodable value is 50 bits. 394 */ 395 396 #define MANTSIZE2 20 /* 20 bit mantissa. */ 397 #define EXPSIZE2 5 /* 5 bit base 2 exponent. */ 398 #define MAXFRACT2 ((1ul << MANTSIZE2) - 1) /* Maximum fractional value. */ 399 #define MAXEXP2 ((1 <<EXPSIZE2) - 1) /* Maximum exponent. */ 400 401 static comp2_t encode_comp2_t(u64 value) 402 { 403 int exp, rnd; 404 405 exp = (value > (MAXFRACT2>>1)); 406 rnd = 0; 407 while (value > MAXFRACT2) { 408 rnd = value & 1; 409 value >>= 1; 410 exp++; 411 } 412 413 /* 414 * If we need to round up, do it (and handle overflow correctly). 415 */ 416 if (rnd && (++value > MAXFRACT2)) { 417 value >>= 1; 418 exp++; 419 } 420 421 if (exp > MAXEXP2) { 422 /* Overflow. Return largest representable number instead. */ 423 return (1ul << (MANTSIZE2+EXPSIZE2-1)) - 1; 424 } else { 425 return (value & (MAXFRACT2>>1)) | (exp << (MANTSIZE2-1)); 426 } 427 } 428 #endif 429 430 #if ACCT_VERSION==3 431 /* 432 * encode an u64 into a 32 bit IEEE float 433 */ 434 static u32 encode_float(u64 value) 435 { 436 unsigned exp = 190; 437 unsigned u; 438 439 if (value==0) return 0; 440 while ((s64)value > 0){ 441 value <<= 1; 442 exp--; 443 } 444 u = (u32)(value >> 40) & 0x7fffffu; 445 return u | (exp << 23); 446 } 447 #endif 448 449 /* 450 * Write an accounting entry for an exiting process 451 * 452 * The acct_process() call is the workhorse of the process 453 * accounting system. The struct acct is built here and then written 454 * into the accounting file. This function should only be called from 455 * do_exit() or when switching to a different output file. 456 */ 457 458 static void fill_ac(acct_t *ac) 459 { 460 struct pacct_struct *pacct = ¤t->signal->pacct; 461 u64 elapsed, run_time; 462 struct tty_struct *tty; 463 464 /* 465 * Fill the accounting struct with the needed info as recorded 466 * by the different kernel functions. 467 */ 468 memset(ac, 0, sizeof(acct_t)); 469 470 ac->ac_version = ACCT_VERSION | ACCT_BYTEORDER; 471 strlcpy(ac->ac_comm, current->comm, sizeof(ac->ac_comm)); 472 473 /* calculate run_time in nsec*/ 474 run_time = ktime_get_ns(); 475 run_time -= current->group_leader->start_time; 476 /* convert nsec -> AHZ */ 477 elapsed = nsec_to_AHZ(run_time); 478 #if ACCT_VERSION==3 479 ac->ac_etime = encode_float(elapsed); 480 #else 481 ac->ac_etime = encode_comp_t(elapsed < (unsigned long) -1l ? 482 (unsigned long) elapsed : (unsigned long) -1l); 483 #endif 484 #if ACCT_VERSION==1 || ACCT_VERSION==2 485 { 486 /* new enlarged etime field */ 487 comp2_t etime = encode_comp2_t(elapsed); 488 ac->ac_etime_hi = etime >> 16; 489 ac->ac_etime_lo = (u16) etime; 490 } 491 #endif 492 do_div(elapsed, AHZ); 493 ac->ac_btime = get_seconds() - elapsed; 494 #if ACCT_VERSION==2 495 ac->ac_ahz = AHZ; 496 #endif 497 498 spin_lock_irq(¤t->sighand->siglock); 499 tty = current->signal->tty; /* Safe as we hold the siglock */ 500 ac->ac_tty = tty ? old_encode_dev(tty_devnum(tty)) : 0; 501 ac->ac_utime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_utime))); 502 ac->ac_stime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_stime))); 503 ac->ac_flag = pacct->ac_flag; 504 ac->ac_mem = encode_comp_t(pacct->ac_mem); 505 ac->ac_minflt = encode_comp_t(pacct->ac_minflt); 506 ac->ac_majflt = encode_comp_t(pacct->ac_majflt); 507 ac->ac_exitcode = pacct->ac_exitcode; 508 spin_unlock_irq(¤t->sighand->siglock); 509 } 510 /* 511 * do_acct_process does all actual work. Caller holds the reference to file. 512 */ 513 static void do_acct_process(struct bsd_acct_struct *acct) 514 { 515 acct_t ac; 516 unsigned long flim; 517 const struct cred *orig_cred; 518 struct pid_namespace *ns = acct->ns; 519 struct file *file = acct->file; 520 521 /* 522 * Accounting records are not subject to resource limits. 523 */ 524 flim = current->signal->rlim[RLIMIT_FSIZE].rlim_cur; 525 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY; 526 /* Perform file operations on behalf of whoever enabled accounting */ 527 orig_cred = override_creds(file->f_cred); 528 529 /* 530 * First check to see if there is enough free_space to continue 531 * the process accounting system. 532 */ 533 if (!check_free_space(acct)) 534 goto out; 535 536 fill_ac(&ac); 537 /* we really need to bite the bullet and change layout */ 538 ac.ac_uid = from_kuid_munged(file->f_cred->user_ns, orig_cred->uid); 539 ac.ac_gid = from_kgid_munged(file->f_cred->user_ns, orig_cred->gid); 540 #if ACCT_VERSION==1 || ACCT_VERSION==2 541 /* backward-compatible 16 bit fields */ 542 ac.ac_uid16 = ac.ac_uid; 543 ac.ac_gid16 = ac.ac_gid; 544 #endif 545 #if ACCT_VERSION==3 546 ac.ac_pid = task_tgid_nr_ns(current, ns); 547 rcu_read_lock(); 548 ac.ac_ppid = task_tgid_nr_ns(rcu_dereference(current->real_parent), ns); 549 rcu_read_unlock(); 550 #endif 551 /* 552 * Get freeze protection. If the fs is frozen, just skip the write 553 * as we could deadlock the system otherwise. 554 */ 555 if (file_start_write_trylock(file)) { 556 /* it's been opened O_APPEND, so position is irrelevant */ 557 loff_t pos = 0; 558 __kernel_write(file, (char *)&ac, sizeof(acct_t), &pos); 559 file_end_write(file); 560 } 561 out: 562 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = flim; 563 revert_creds(orig_cred); 564 } 565 566 /** 567 * acct_collect - collect accounting information into pacct_struct 568 * @exitcode: task exit code 569 * @group_dead: not 0, if this thread is the last one in the process. 570 */ 571 void acct_collect(long exitcode, int group_dead) 572 { 573 struct pacct_struct *pacct = ¤t->signal->pacct; 574 cputime_t utime, stime; 575 unsigned long vsize = 0; 576 577 if (group_dead && current->mm) { 578 struct vm_area_struct *vma; 579 down_read(¤t->mm->mmap_sem); 580 vma = current->mm->mmap; 581 while (vma) { 582 vsize += vma->vm_end - vma->vm_start; 583 vma = vma->vm_next; 584 } 585 up_read(¤t->mm->mmap_sem); 586 } 587 588 spin_lock_irq(¤t->sighand->siglock); 589 if (group_dead) 590 pacct->ac_mem = vsize / 1024; 591 if (thread_group_leader(current)) { 592 pacct->ac_exitcode = exitcode; 593 if (current->flags & PF_FORKNOEXEC) 594 pacct->ac_flag |= AFORK; 595 } 596 if (current->flags & PF_SUPERPRIV) 597 pacct->ac_flag |= ASU; 598 if (current->flags & PF_DUMPCORE) 599 pacct->ac_flag |= ACORE; 600 if (current->flags & PF_SIGNALED) 601 pacct->ac_flag |= AXSIG; 602 task_cputime(current, &utime, &stime); 603 pacct->ac_utime += utime; 604 pacct->ac_stime += stime; 605 pacct->ac_minflt += current->min_flt; 606 pacct->ac_majflt += current->maj_flt; 607 spin_unlock_irq(¤t->sighand->siglock); 608 } 609 610 static void slow_acct_process(struct pid_namespace *ns) 611 { 612 for ( ; ns; ns = ns->parent) { 613 struct bsd_acct_struct *acct = acct_get(ns); 614 if (acct) { 615 do_acct_process(acct); 616 mutex_unlock(&acct->lock); 617 acct_put(acct); 618 } 619 } 620 } 621 622 /** 623 * acct_process 624 * 625 * handles process accounting for an exiting task 626 */ 627 void acct_process(void) 628 { 629 struct pid_namespace *ns; 630 631 /* 632 * This loop is safe lockless, since current is still 633 * alive and holds its namespace, which in turn holds 634 * its parent. 635 */ 636 for (ns = task_active_pid_ns(current); ns != NULL; ns = ns->parent) { 637 if (ns->bacct) 638 break; 639 } 640 if (unlikely(ns)) 641 slow_acct_process(ns); 642 } 643