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