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