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 <linux/fs_pin.h> 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 struct fs_pin pin; 83 struct mutex lock; 84 int active; 85 unsigned long needcheck; 86 struct file *file; 87 struct pid_namespace *ns; 88 struct work_struct work; 89 struct completion done; 90 }; 91 92 /* 93 * Check the amount of free space and suspend/resume accordingly. 94 */ 95 static int check_free_space(struct bsd_acct_struct *acct) 96 { 97 struct kstatfs sbuf; 98 99 if (time_is_before_jiffies(acct->needcheck)) 100 goto out; 101 102 /* May block */ 103 if (vfs_statfs(&acct->file->f_path, &sbuf)) 104 goto out; 105 106 if (acct->active) { 107 u64 suspend = sbuf.f_blocks * SUSPEND; 108 do_div(suspend, 100); 109 if (sbuf.f_bavail <= suspend) { 110 acct->active = 0; 111 pr_info("Process accounting paused\n"); 112 } 113 } else { 114 u64 resume = sbuf.f_blocks * RESUME; 115 do_div(resume, 100); 116 if (sbuf.f_bavail >= resume) { 117 acct->active = 1; 118 pr_info("Process accounting resumed\n"); 119 } 120 } 121 122 acct->needcheck = jiffies + ACCT_TIMEOUT*HZ; 123 out: 124 return acct->active; 125 } 126 127 static void acct_put(struct bsd_acct_struct *p) 128 { 129 if (atomic_long_dec_and_test(&p->pin.count)) 130 kfree_rcu(p, pin.rcu); 131 } 132 133 static struct bsd_acct_struct *acct_get(struct pid_namespace *ns) 134 { 135 struct bsd_acct_struct *res; 136 again: 137 smp_rmb(); 138 rcu_read_lock(); 139 res = ACCESS_ONCE(ns->bacct); 140 if (!res) { 141 rcu_read_unlock(); 142 return NULL; 143 } 144 if (!atomic_long_inc_not_zero(&res->pin.count)) { 145 rcu_read_unlock(); 146 cpu_relax(); 147 goto again; 148 } 149 rcu_read_unlock(); 150 mutex_lock(&res->lock); 151 if (!res->ns) { 152 mutex_unlock(&res->lock); 153 acct_put(res); 154 goto again; 155 } 156 return res; 157 } 158 159 static void close_work(struct work_struct *work) 160 { 161 struct bsd_acct_struct *acct = container_of(work, struct bsd_acct_struct, work); 162 struct file *file = acct->file; 163 if (file->f_op->flush) 164 file->f_op->flush(file, NULL); 165 __fput_sync(file); 166 complete(&acct->done); 167 } 168 169 static void acct_kill(struct bsd_acct_struct *acct, 170 struct bsd_acct_struct *new) 171 { 172 if (acct) { 173 struct pid_namespace *ns = acct->ns; 174 do_acct_process(acct); 175 INIT_WORK(&acct->work, close_work); 176 init_completion(&acct->done); 177 schedule_work(&acct->work); 178 wait_for_completion(&acct->done); 179 pin_remove(&acct->pin); 180 ns->bacct = new; 181 acct->ns = NULL; 182 atomic_long_dec(&acct->pin.count); 183 mutex_unlock(&acct->lock); 184 acct_put(acct); 185 } 186 } 187 188 static void acct_pin_kill(struct fs_pin *pin) 189 { 190 struct bsd_acct_struct *acct; 191 acct = container_of(pin, struct bsd_acct_struct, pin); 192 if (!atomic_long_inc_not_zero(&pin->count)) { 193 rcu_read_unlock(); 194 cpu_relax(); 195 return; 196 } 197 rcu_read_unlock(); 198 mutex_lock(&acct->lock); 199 if (!acct->ns) { 200 mutex_unlock(&acct->lock); 201 acct_put(acct); 202 acct = NULL; 203 } 204 acct_kill(acct, NULL); 205 } 206 207 static int acct_on(struct filename *pathname) 208 { 209 struct file *file; 210 struct vfsmount *mnt, *internal; 211 struct pid_namespace *ns = task_active_pid_ns(current); 212 struct bsd_acct_struct *acct, *old; 213 int err; 214 215 acct = kzalloc(sizeof(struct bsd_acct_struct), GFP_KERNEL); 216 if (!acct) 217 return -ENOMEM; 218 219 /* Difference from BSD - they don't do O_APPEND */ 220 file = file_open_name(pathname, O_WRONLY|O_APPEND|O_LARGEFILE, 0); 221 if (IS_ERR(file)) { 222 kfree(acct); 223 return PTR_ERR(file); 224 } 225 226 if (!S_ISREG(file_inode(file)->i_mode)) { 227 kfree(acct); 228 filp_close(file, NULL); 229 return -EACCES; 230 } 231 232 if (!file->f_op->write) { 233 kfree(acct); 234 filp_close(file, NULL); 235 return -EIO; 236 } 237 internal = mnt_clone_internal(&file->f_path); 238 if (IS_ERR(internal)) { 239 kfree(acct); 240 filp_close(file, NULL); 241 return PTR_ERR(internal); 242 } 243 err = mnt_want_write(internal); 244 if (err) { 245 mntput(internal); 246 kfree(acct); 247 filp_close(file, NULL); 248 return err; 249 } 250 mnt = file->f_path.mnt; 251 file->f_path.mnt = internal; 252 253 atomic_long_set(&acct->pin.count, 1); 254 acct->pin.kill = acct_pin_kill; 255 acct->file = file; 256 acct->needcheck = jiffies; 257 acct->ns = ns; 258 mutex_init(&acct->lock); 259 mutex_lock_nested(&acct->lock, 1); /* nobody has seen it yet */ 260 pin_insert(&acct->pin, mnt); 261 262 old = acct_get(ns); 263 if (old) 264 acct_kill(old, acct); 265 else 266 ns->bacct = acct; 267 mutex_unlock(&acct->lock); 268 mnt_drop_write(mnt); 269 mntput(mnt); 270 return 0; 271 } 272 273 static DEFINE_MUTEX(acct_on_mutex); 274 275 /** 276 * sys_acct - enable/disable process accounting 277 * @name: file name for accounting records or NULL to shutdown accounting 278 * 279 * Returns 0 for success or negative errno values for failure. 280 * 281 * sys_acct() is the only system call needed to implement process 282 * accounting. It takes the name of the file where accounting records 283 * should be written. If the filename is NULL, accounting will be 284 * shutdown. 285 */ 286 SYSCALL_DEFINE1(acct, const char __user *, name) 287 { 288 int error = 0; 289 290 if (!capable(CAP_SYS_PACCT)) 291 return -EPERM; 292 293 if (name) { 294 struct filename *tmp = getname(name); 295 296 if (IS_ERR(tmp)) 297 return PTR_ERR(tmp); 298 mutex_lock(&acct_on_mutex); 299 error = acct_on(tmp); 300 mutex_unlock(&acct_on_mutex); 301 putname(tmp); 302 } else { 303 acct_kill(acct_get(task_active_pid_ns(current)), NULL); 304 } 305 306 return error; 307 } 308 309 void acct_exit_ns(struct pid_namespace *ns) 310 { 311 acct_kill(acct_get(ns), NULL); 312 } 313 314 /* 315 * encode an unsigned long into a comp_t 316 * 317 * This routine has been adopted from the encode_comp_t() function in 318 * the kern_acct.c file of the FreeBSD operating system. The encoding 319 * is a 13-bit fraction with a 3-bit (base 8) exponent. 320 */ 321 322 #define MANTSIZE 13 /* 13 bit mantissa. */ 323 #define EXPSIZE 3 /* Base 8 (3 bit) exponent. */ 324 #define MAXFRACT ((1 << MANTSIZE) - 1) /* Maximum fractional value. */ 325 326 static comp_t encode_comp_t(unsigned long value) 327 { 328 int exp, rnd; 329 330 exp = rnd = 0; 331 while (value > MAXFRACT) { 332 rnd = value & (1 << (EXPSIZE - 1)); /* Round up? */ 333 value >>= EXPSIZE; /* Base 8 exponent == 3 bit shift. */ 334 exp++; 335 } 336 337 /* 338 * If we need to round up, do it (and handle overflow correctly). 339 */ 340 if (rnd && (++value > MAXFRACT)) { 341 value >>= EXPSIZE; 342 exp++; 343 } 344 345 /* 346 * Clean it up and polish it off. 347 */ 348 exp <<= MANTSIZE; /* Shift the exponent into place */ 349 exp += value; /* and add on the mantissa. */ 350 return exp; 351 } 352 353 #if ACCT_VERSION == 1 || ACCT_VERSION == 2 354 /* 355 * encode an u64 into a comp2_t (24 bits) 356 * 357 * Format: 5 bit base 2 exponent, 20 bits mantissa. 358 * The leading bit of the mantissa is not stored, but implied for 359 * non-zero exponents. 360 * Largest encodable value is 50 bits. 361 */ 362 363 #define MANTSIZE2 20 /* 20 bit mantissa. */ 364 #define EXPSIZE2 5 /* 5 bit base 2 exponent. */ 365 #define MAXFRACT2 ((1ul << MANTSIZE2) - 1) /* Maximum fractional value. */ 366 #define MAXEXP2 ((1 << EXPSIZE2) - 1) /* Maximum exponent. */ 367 368 static comp2_t encode_comp2_t(u64 value) 369 { 370 int exp, rnd; 371 372 exp = (value > (MAXFRACT2>>1)); 373 rnd = 0; 374 while (value > MAXFRACT2) { 375 rnd = value & 1; 376 value >>= 1; 377 exp++; 378 } 379 380 /* 381 * If we need to round up, do it (and handle overflow correctly). 382 */ 383 if (rnd && (++value > MAXFRACT2)) { 384 value >>= 1; 385 exp++; 386 } 387 388 if (exp > MAXEXP2) { 389 /* Overflow. Return largest representable number instead. */ 390 return (1ul << (MANTSIZE2+EXPSIZE2-1)) - 1; 391 } else { 392 return (value & (MAXFRACT2>>1)) | (exp << (MANTSIZE2-1)); 393 } 394 } 395 #endif 396 397 #if ACCT_VERSION == 3 398 /* 399 * encode an u64 into a 32 bit IEEE float 400 */ 401 static u32 encode_float(u64 value) 402 { 403 unsigned exp = 190; 404 unsigned u; 405 406 if (value == 0) 407 return 0; 408 while ((s64)value > 0) { 409 value <<= 1; 410 exp--; 411 } 412 u = (u32)(value >> 40) & 0x7fffffu; 413 return u | (exp << 23); 414 } 415 #endif 416 417 /* 418 * Write an accounting entry for an exiting process 419 * 420 * The acct_process() call is the workhorse of the process 421 * accounting system. The struct acct is built here and then written 422 * into the accounting file. This function should only be called from 423 * do_exit() or when switching to a different output file. 424 */ 425 426 static void fill_ac(acct_t *ac) 427 { 428 struct pacct_struct *pacct = ¤t->signal->pacct; 429 u64 elapsed, run_time; 430 struct tty_struct *tty; 431 432 /* 433 * Fill the accounting struct with the needed info as recorded 434 * by the different kernel functions. 435 */ 436 memset(ac, 0, sizeof(acct_t)); 437 438 ac->ac_version = ACCT_VERSION | ACCT_BYTEORDER; 439 strlcpy(ac->ac_comm, current->comm, sizeof(ac->ac_comm)); 440 441 /* calculate run_time in nsec*/ 442 run_time = ktime_get_ns(); 443 run_time -= current->group_leader->start_time; 444 /* convert nsec -> AHZ */ 445 elapsed = nsec_to_AHZ(run_time); 446 #if ACCT_VERSION == 3 447 ac->ac_etime = encode_float(elapsed); 448 #else 449 ac->ac_etime = encode_comp_t(elapsed < (unsigned long) -1l ? 450 (unsigned long) elapsed : (unsigned long) -1l); 451 #endif 452 #if ACCT_VERSION == 1 || ACCT_VERSION == 2 453 { 454 /* new enlarged etime field */ 455 comp2_t etime = encode_comp2_t(elapsed); 456 457 ac->ac_etime_hi = etime >> 16; 458 ac->ac_etime_lo = (u16) etime; 459 } 460 #endif 461 do_div(elapsed, AHZ); 462 ac->ac_btime = get_seconds() - elapsed; 463 #if ACCT_VERSION==2 464 ac->ac_ahz = AHZ; 465 #endif 466 467 spin_lock_irq(¤t->sighand->siglock); 468 tty = current->signal->tty; /* Safe as we hold the siglock */ 469 ac->ac_tty = tty ? old_encode_dev(tty_devnum(tty)) : 0; 470 ac->ac_utime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_utime))); 471 ac->ac_stime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_stime))); 472 ac->ac_flag = pacct->ac_flag; 473 ac->ac_mem = encode_comp_t(pacct->ac_mem); 474 ac->ac_minflt = encode_comp_t(pacct->ac_minflt); 475 ac->ac_majflt = encode_comp_t(pacct->ac_majflt); 476 ac->ac_exitcode = pacct->ac_exitcode; 477 spin_unlock_irq(¤t->sighand->siglock); 478 } 479 /* 480 * do_acct_process does all actual work. Caller holds the reference to file. 481 */ 482 static void do_acct_process(struct bsd_acct_struct *acct) 483 { 484 acct_t ac; 485 unsigned long flim; 486 const struct cred *orig_cred; 487 struct file *file = acct->file; 488 489 /* 490 * Accounting records are not subject to resource limits. 491 */ 492 flim = current->signal->rlim[RLIMIT_FSIZE].rlim_cur; 493 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY; 494 /* Perform file operations on behalf of whoever enabled accounting */ 495 orig_cred = override_creds(file->f_cred); 496 497 /* 498 * First check to see if there is enough free_space to continue 499 * the process accounting system. 500 */ 501 if (!check_free_space(acct)) 502 goto out; 503 504 fill_ac(&ac); 505 /* we really need to bite the bullet and change layout */ 506 ac.ac_uid = from_kuid_munged(file->f_cred->user_ns, orig_cred->uid); 507 ac.ac_gid = from_kgid_munged(file->f_cred->user_ns, orig_cred->gid); 508 #if ACCT_VERSION == 1 || ACCT_VERSION == 2 509 /* backward-compatible 16 bit fields */ 510 ac.ac_uid16 = ac.ac_uid; 511 ac.ac_gid16 = ac.ac_gid; 512 #endif 513 #if ACCT_VERSION == 3 514 { 515 struct pid_namespace *ns = acct->ns; 516 517 ac.ac_pid = task_tgid_nr_ns(current, ns); 518 rcu_read_lock(); 519 ac.ac_ppid = task_tgid_nr_ns(rcu_dereference(current->real_parent), 520 ns); 521 rcu_read_unlock(); 522 } 523 #endif 524 /* 525 * Get freeze protection. If the fs is frozen, just skip the write 526 * as we could deadlock the system otherwise. 527 */ 528 if (file_start_write_trylock(file)) { 529 /* it's been opened O_APPEND, so position is irrelevant */ 530 loff_t pos = 0; 531 __kernel_write(file, (char *)&ac, sizeof(acct_t), &pos); 532 file_end_write(file); 533 } 534 out: 535 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = flim; 536 revert_creds(orig_cred); 537 } 538 539 /** 540 * acct_collect - collect accounting information into pacct_struct 541 * @exitcode: task exit code 542 * @group_dead: not 0, if this thread is the last one in the process. 543 */ 544 void acct_collect(long exitcode, int group_dead) 545 { 546 struct pacct_struct *pacct = ¤t->signal->pacct; 547 cputime_t utime, stime; 548 unsigned long vsize = 0; 549 550 if (group_dead && current->mm) { 551 struct vm_area_struct *vma; 552 553 down_read(¤t->mm->mmap_sem); 554 vma = current->mm->mmap; 555 while (vma) { 556 vsize += vma->vm_end - vma->vm_start; 557 vma = vma->vm_next; 558 } 559 up_read(¤t->mm->mmap_sem); 560 } 561 562 spin_lock_irq(¤t->sighand->siglock); 563 if (group_dead) 564 pacct->ac_mem = vsize / 1024; 565 if (thread_group_leader(current)) { 566 pacct->ac_exitcode = exitcode; 567 if (current->flags & PF_FORKNOEXEC) 568 pacct->ac_flag |= AFORK; 569 } 570 if (current->flags & PF_SUPERPRIV) 571 pacct->ac_flag |= ASU; 572 if (current->flags & PF_DUMPCORE) 573 pacct->ac_flag |= ACORE; 574 if (current->flags & PF_SIGNALED) 575 pacct->ac_flag |= AXSIG; 576 task_cputime(current, &utime, &stime); 577 pacct->ac_utime += utime; 578 pacct->ac_stime += stime; 579 pacct->ac_minflt += current->min_flt; 580 pacct->ac_majflt += current->maj_flt; 581 spin_unlock_irq(¤t->sighand->siglock); 582 } 583 584 static void slow_acct_process(struct pid_namespace *ns) 585 { 586 for ( ; ns; ns = ns->parent) { 587 struct bsd_acct_struct *acct = acct_get(ns); 588 if (acct) { 589 do_acct_process(acct); 590 mutex_unlock(&acct->lock); 591 acct_put(acct); 592 } 593 } 594 } 595 596 /** 597 * acct_process 598 * 599 * handles process accounting for an exiting task 600 */ 601 void acct_process(void) 602 { 603 struct pid_namespace *ns; 604 605 /* 606 * This loop is safe lockless, since current is still 607 * alive and holds its namespace, which in turn holds 608 * its parent. 609 */ 610 for (ns = task_active_pid_ns(current); ns != NULL; ns = ns->parent) { 611 if (ns->bacct) 612 break; 613 } 614 if (unlikely(ns)) 615 slow_acct_process(ns); 616 } 617