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