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 spin_lock(&acct_lock); 210 hlist_add_head(&new->s_list, &m->mnt_sb->s_pins); 211 hlist_add_head(&new->m_list, &real_mount(m)->mnt_pins); 212 spin_unlock(&acct_lock); 213 mutex_unlock(&new->lock); 214 } 215 acct->ns = NULL; 216 atomic_long_dec(&acct->count); 217 mutex_unlock(&acct->lock); 218 acct_put(acct); 219 } 220 } 221 222 static int acct_on(struct filename *pathname) 223 { 224 struct file *file; 225 struct vfsmount *mnt; 226 struct pid_namespace *ns = task_active_pid_ns(current); 227 struct bsd_acct_struct *acct, *old; 228 229 acct = kzalloc(sizeof(struct bsd_acct_struct), GFP_KERNEL); 230 if (!acct) 231 return -ENOMEM; 232 233 /* Difference from BSD - they don't do O_APPEND */ 234 file = file_open_name(pathname, O_WRONLY|O_APPEND|O_LARGEFILE, 0); 235 if (IS_ERR(file)) { 236 kfree(acct); 237 return PTR_ERR(file); 238 } 239 240 if (!S_ISREG(file_inode(file)->i_mode)) { 241 kfree(acct); 242 filp_close(file, NULL); 243 return -EACCES; 244 } 245 246 if (!file->f_op->write) { 247 kfree(acct); 248 filp_close(file, NULL); 249 return -EIO; 250 } 251 252 atomic_long_set(&acct->count, 1); 253 acct->file = file; 254 acct->needcheck = jiffies; 255 acct->ns = ns; 256 mutex_init(&acct->lock); 257 mnt = file->f_path.mnt; 258 mnt_pin(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 hlist_add_head(&acct->s_list, &mnt->mnt_sb->s_pins); 268 hlist_add_head(&acct->m_list, &real_mount(mnt)->mnt_pins); 269 spin_unlock(&acct_lock); 270 mutex_unlock(&acct->lock); 271 } 272 mntput(mnt); /* it's pinned, now give up active reference */ 273 return 0; 274 } 275 276 static DEFINE_MUTEX(acct_on_mutex); 277 278 /** 279 * sys_acct - enable/disable process accounting 280 * @name: file name for accounting records or NULL to shutdown accounting 281 * 282 * Returns 0 for success or negative errno values for failure. 283 * 284 * sys_acct() is the only system call needed to implement process 285 * accounting. It takes the name of the file where accounting records 286 * should be written. If the filename is NULL, accounting will be 287 * shutdown. 288 */ 289 SYSCALL_DEFINE1(acct, const char __user *, name) 290 { 291 int error = 0; 292 293 if (!capable(CAP_SYS_PACCT)) 294 return -EPERM; 295 296 if (name) { 297 struct filename *tmp = getname(name); 298 if (IS_ERR(tmp)) 299 return PTR_ERR(tmp); 300 mutex_lock(&acct_on_mutex); 301 error = acct_on(tmp); 302 mutex_unlock(&acct_on_mutex); 303 putname(tmp); 304 } else { 305 acct_kill(acct_get(task_active_pid_ns(current)), NULL); 306 } 307 308 return error; 309 } 310 311 void acct_auto_close_mnt(struct hlist_head *list) 312 { 313 rcu_read_lock(); 314 while (1) { 315 struct hlist_node *p = ACCESS_ONCE(list->first); 316 if (!p) 317 break; 318 acct_kill(__acct_get(hlist_entry(p, 319 struct bsd_acct_struct, 320 m_list)), NULL); 321 rcu_read_lock(); 322 } 323 rcu_read_unlock(); 324 } 325 326 void acct_auto_close(struct hlist_head *list) 327 { 328 rcu_read_lock(); 329 while (1) { 330 struct hlist_node *p = ACCESS_ONCE(list->first); 331 if (!p) 332 break; 333 acct_kill(__acct_get(hlist_entry(p, 334 struct bsd_acct_struct, 335 s_list)), NULL); 336 rcu_read_lock(); 337 } 338 rcu_read_unlock(); 339 } 340 341 void acct_exit_ns(struct pid_namespace *ns) 342 { 343 acct_kill(acct_get(ns), NULL); 344 } 345 346 /* 347 * encode an unsigned long into a comp_t 348 * 349 * This routine has been adopted from the encode_comp_t() function in 350 * the kern_acct.c file of the FreeBSD operating system. The encoding 351 * is a 13-bit fraction with a 3-bit (base 8) exponent. 352 */ 353 354 #define MANTSIZE 13 /* 13 bit mantissa. */ 355 #define EXPSIZE 3 /* Base 8 (3 bit) exponent. */ 356 #define MAXFRACT ((1 << MANTSIZE) - 1) /* Maximum fractional value. */ 357 358 static comp_t encode_comp_t(unsigned long value) 359 { 360 int exp, rnd; 361 362 exp = rnd = 0; 363 while (value > MAXFRACT) { 364 rnd = value & (1 << (EXPSIZE - 1)); /* Round up? */ 365 value >>= EXPSIZE; /* Base 8 exponent == 3 bit shift. */ 366 exp++; 367 } 368 369 /* 370 * If we need to round up, do it (and handle overflow correctly). 371 */ 372 if (rnd && (++value > MAXFRACT)) { 373 value >>= EXPSIZE; 374 exp++; 375 } 376 377 /* 378 * Clean it up and polish it off. 379 */ 380 exp <<= MANTSIZE; /* Shift the exponent into place */ 381 exp += value; /* and add on the mantissa. */ 382 return exp; 383 } 384 385 #if ACCT_VERSION==1 || ACCT_VERSION==2 386 /* 387 * encode an u64 into a comp2_t (24 bits) 388 * 389 * Format: 5 bit base 2 exponent, 20 bits mantissa. 390 * The leading bit of the mantissa is not stored, but implied for 391 * non-zero exponents. 392 * Largest encodable value is 50 bits. 393 */ 394 395 #define MANTSIZE2 20 /* 20 bit mantissa. */ 396 #define EXPSIZE2 5 /* 5 bit base 2 exponent. */ 397 #define MAXFRACT2 ((1ul << MANTSIZE2) - 1) /* Maximum fractional value. */ 398 #define MAXEXP2 ((1 <<EXPSIZE2) - 1) /* Maximum exponent. */ 399 400 static comp2_t encode_comp2_t(u64 value) 401 { 402 int exp, rnd; 403 404 exp = (value > (MAXFRACT2>>1)); 405 rnd = 0; 406 while (value > MAXFRACT2) { 407 rnd = value & 1; 408 value >>= 1; 409 exp++; 410 } 411 412 /* 413 * If we need to round up, do it (and handle overflow correctly). 414 */ 415 if (rnd && (++value > MAXFRACT2)) { 416 value >>= 1; 417 exp++; 418 } 419 420 if (exp > MAXEXP2) { 421 /* Overflow. Return largest representable number instead. */ 422 return (1ul << (MANTSIZE2+EXPSIZE2-1)) - 1; 423 } else { 424 return (value & (MAXFRACT2>>1)) | (exp << (MANTSIZE2-1)); 425 } 426 } 427 #endif 428 429 #if ACCT_VERSION==3 430 /* 431 * encode an u64 into a 32 bit IEEE float 432 */ 433 static u32 encode_float(u64 value) 434 { 435 unsigned exp = 190; 436 unsigned u; 437 438 if (value==0) return 0; 439 while ((s64)value > 0){ 440 value <<= 1; 441 exp--; 442 } 443 u = (u32)(value >> 40) & 0x7fffffu; 444 return u | (exp << 23); 445 } 446 #endif 447 448 /* 449 * Write an accounting entry for an exiting process 450 * 451 * The acct_process() call is the workhorse of the process 452 * accounting system. The struct acct is built here and then written 453 * into the accounting file. This function should only be called from 454 * do_exit() or when switching to a different output file. 455 */ 456 457 static void fill_ac(acct_t *ac) 458 { 459 struct pacct_struct *pacct = ¤t->signal->pacct; 460 u64 elapsed, run_time; 461 struct tty_struct *tty; 462 463 /* 464 * Fill the accounting struct with the needed info as recorded 465 * by the different kernel functions. 466 */ 467 memset(ac, 0, sizeof(acct_t)); 468 469 ac->ac_version = ACCT_VERSION | ACCT_BYTEORDER; 470 strlcpy(ac->ac_comm, current->comm, sizeof(ac->ac_comm)); 471 472 /* calculate run_time in nsec*/ 473 run_time = ktime_get_ns(); 474 run_time -= current->group_leader->start_time; 475 /* convert nsec -> AHZ */ 476 elapsed = nsec_to_AHZ(run_time); 477 #if ACCT_VERSION==3 478 ac->ac_etime = encode_float(elapsed); 479 #else 480 ac->ac_etime = encode_comp_t(elapsed < (unsigned long) -1l ? 481 (unsigned long) elapsed : (unsigned long) -1l); 482 #endif 483 #if ACCT_VERSION==1 || ACCT_VERSION==2 484 { 485 /* new enlarged etime field */ 486 comp2_t etime = encode_comp2_t(elapsed); 487 ac->ac_etime_hi = etime >> 16; 488 ac->ac_etime_lo = (u16) etime; 489 } 490 #endif 491 do_div(elapsed, AHZ); 492 ac->ac_btime = get_seconds() - elapsed; 493 #if ACCT_VERSION==2 494 ac->ac_ahz = AHZ; 495 #endif 496 497 spin_lock_irq(¤t->sighand->siglock); 498 tty = current->signal->tty; /* Safe as we hold the siglock */ 499 ac->ac_tty = tty ? old_encode_dev(tty_devnum(tty)) : 0; 500 ac->ac_utime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_utime))); 501 ac->ac_stime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_stime))); 502 ac->ac_flag = pacct->ac_flag; 503 ac->ac_mem = encode_comp_t(pacct->ac_mem); 504 ac->ac_minflt = encode_comp_t(pacct->ac_minflt); 505 ac->ac_majflt = encode_comp_t(pacct->ac_majflt); 506 ac->ac_exitcode = pacct->ac_exitcode; 507 spin_unlock_irq(¤t->sighand->siglock); 508 } 509 /* 510 * do_acct_process does all actual work. Caller holds the reference to file. 511 */ 512 static void do_acct_process(struct bsd_acct_struct *acct) 513 { 514 acct_t ac; 515 unsigned long flim; 516 const struct cred *orig_cred; 517 struct pid_namespace *ns = acct->ns; 518 struct file *file = acct->file; 519 520 /* 521 * Accounting records are not subject to resource limits. 522 */ 523 flim = current->signal->rlim[RLIMIT_FSIZE].rlim_cur; 524 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY; 525 /* Perform file operations on behalf of whoever enabled accounting */ 526 orig_cred = override_creds(file->f_cred); 527 528 /* 529 * First check to see if there is enough free_space to continue 530 * the process accounting system. 531 */ 532 if (!check_free_space(acct)) 533 goto out; 534 535 fill_ac(&ac); 536 /* we really need to bite the bullet and change layout */ 537 ac.ac_uid = from_kuid_munged(file->f_cred->user_ns, orig_cred->uid); 538 ac.ac_gid = from_kgid_munged(file->f_cred->user_ns, orig_cred->gid); 539 #if ACCT_VERSION==1 || ACCT_VERSION==2 540 /* backward-compatible 16 bit fields */ 541 ac.ac_uid16 = ac.ac_uid; 542 ac.ac_gid16 = ac.ac_gid; 543 #endif 544 #if ACCT_VERSION==3 545 ac.ac_pid = task_tgid_nr_ns(current, ns); 546 rcu_read_lock(); 547 ac.ac_ppid = task_tgid_nr_ns(rcu_dereference(current->real_parent), ns); 548 rcu_read_unlock(); 549 #endif 550 /* 551 * Get freeze protection. If the fs is frozen, just skip the write 552 * as we could deadlock the system otherwise. 553 */ 554 if (file_start_write_trylock(file)) { 555 /* it's been opened O_APPEND, so position is irrelevant */ 556 loff_t pos = 0; 557 __kernel_write(file, (char *)&ac, sizeof(acct_t), &pos); 558 file_end_write(file); 559 } 560 out: 561 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = flim; 562 revert_creds(orig_cred); 563 } 564 565 /** 566 * acct_collect - collect accounting information into pacct_struct 567 * @exitcode: task exit code 568 * @group_dead: not 0, if this thread is the last one in the process. 569 */ 570 void acct_collect(long exitcode, int group_dead) 571 { 572 struct pacct_struct *pacct = ¤t->signal->pacct; 573 cputime_t utime, stime; 574 unsigned long vsize = 0; 575 576 if (group_dead && current->mm) { 577 struct vm_area_struct *vma; 578 down_read(¤t->mm->mmap_sem); 579 vma = current->mm->mmap; 580 while (vma) { 581 vsize += vma->vm_end - vma->vm_start; 582 vma = vma->vm_next; 583 } 584 up_read(¤t->mm->mmap_sem); 585 } 586 587 spin_lock_irq(¤t->sighand->siglock); 588 if (group_dead) 589 pacct->ac_mem = vsize / 1024; 590 if (thread_group_leader(current)) { 591 pacct->ac_exitcode = exitcode; 592 if (current->flags & PF_FORKNOEXEC) 593 pacct->ac_flag |= AFORK; 594 } 595 if (current->flags & PF_SUPERPRIV) 596 pacct->ac_flag |= ASU; 597 if (current->flags & PF_DUMPCORE) 598 pacct->ac_flag |= ACORE; 599 if (current->flags & PF_SIGNALED) 600 pacct->ac_flag |= AXSIG; 601 task_cputime(current, &utime, &stime); 602 pacct->ac_utime += utime; 603 pacct->ac_stime += stime; 604 pacct->ac_minflt += current->min_flt; 605 pacct->ac_majflt += current->maj_flt; 606 spin_unlock_irq(¤t->sighand->siglock); 607 } 608 609 static void slow_acct_process(struct pid_namespace *ns) 610 { 611 for ( ; ns; ns = ns->parent) { 612 struct bsd_acct_struct *acct = acct_get(ns); 613 if (acct) { 614 do_acct_process(acct); 615 mutex_unlock(&acct->lock); 616 acct_put(acct); 617 } 618 } 619 } 620 621 /** 622 * acct_process 623 * 624 * handles process accounting for an exiting task 625 */ 626 void acct_process(void) 627 { 628 struct pid_namespace *ns; 629 630 /* 631 * This loop is safe lockless, since current is still 632 * alive and holds its namespace, which in turn holds 633 * its parent. 634 */ 635 for (ns = task_active_pid_ns(current); ns != NULL; ns = ns->parent) { 636 if (ns->bacct) 637 break; 638 } 639 if (unlikely(ns)) 640 slow_acct_process(ns); 641 } 642