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 long count; 83 struct hlist_node s_list; 84 struct hlist_node m_list; 85 struct mutex lock; 86 int active; 87 unsigned long needcheck; 88 struct file *file; 89 struct pid_namespace *ns; 90 }; 91 92 static DEFINE_SPINLOCK(acct_lock); 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 printk(KERN_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 printk(KERN_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 spin_lock(&acct_lock); 132 if (!--p->count) 133 kfree(p); 134 spin_unlock(&acct_lock); 135 } 136 137 static struct bsd_acct_struct *__acct_get(struct bsd_acct_struct *res) 138 { 139 res->count++; 140 spin_unlock(&acct_lock); 141 mutex_lock(&res->lock); 142 if (!res->ns) { 143 mutex_unlock(&res->lock); 144 spin_lock(&acct_lock); 145 if (!--res->count) 146 kfree(res); 147 return NULL; 148 } 149 return res; 150 } 151 152 static struct bsd_acct_struct *acct_get(struct pid_namespace *ns) 153 { 154 struct bsd_acct_struct *res; 155 spin_lock(&acct_lock); 156 again: 157 if (!ns->bacct) { 158 spin_unlock(&acct_lock); 159 return NULL; 160 } 161 res = __acct_get(ns->bacct); 162 if (!res) 163 goto again; 164 return res; 165 } 166 167 static void acct_kill(struct bsd_acct_struct *acct, 168 struct bsd_acct_struct *new) 169 { 170 if (acct) { 171 struct file *file = acct->file; 172 struct pid_namespace *ns = acct->ns; 173 spin_lock(&acct_lock); 174 hlist_del(&acct->m_list); 175 hlist_del(&acct->s_list); 176 mnt_unpin(file->f_path.mnt); 177 spin_unlock(&acct_lock); 178 do_acct_process(acct); 179 filp_close(file, NULL); 180 spin_lock(&acct_lock); 181 ns->bacct = new; 182 if (new) { 183 struct vfsmount *m = new->file->f_path.mnt; 184 mnt_pin(m); 185 hlist_add_head(&new->s_list, &m->mnt_sb->s_pins); 186 hlist_add_head(&new->m_list, &real_mount(m)->mnt_pins); 187 } 188 acct->ns = NULL; 189 mutex_unlock(&acct->lock); 190 if (!(acct->count -= 2)) 191 kfree(acct); 192 spin_unlock(&acct_lock); 193 } 194 } 195 196 static int acct_on(struct filename *pathname) 197 { 198 struct file *file; 199 struct vfsmount *mnt; 200 struct pid_namespace *ns = task_active_pid_ns(current); 201 struct bsd_acct_struct *acct, *old; 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 } 225 226 acct->count = 1; 227 acct->file = file; 228 acct->needcheck = jiffies; 229 acct->ns = ns; 230 mutex_init(&acct->lock); 231 mnt = file->f_path.mnt; 232 233 old = acct_get(ns); 234 if (old) { 235 acct_kill(old, acct); 236 } else { 237 spin_lock(&acct_lock); 238 ns->bacct = acct; 239 mnt_pin(mnt); 240 hlist_add_head(&acct->s_list, &mnt->mnt_sb->s_pins); 241 hlist_add_head(&acct->m_list, &real_mount(mnt)->mnt_pins); 242 spin_unlock(&acct_lock); 243 } 244 mntput(mnt); /* it's pinned, now give up active reference */ 245 return 0; 246 } 247 248 static DEFINE_MUTEX(acct_on_mutex); 249 250 /** 251 * sys_acct - enable/disable process accounting 252 * @name: file name for accounting records or NULL to shutdown accounting 253 * 254 * Returns 0 for success or negative errno values for failure. 255 * 256 * sys_acct() is the only system call needed to implement process 257 * accounting. It takes the name of the file where accounting records 258 * should be written. If the filename is NULL, accounting will be 259 * shutdown. 260 */ 261 SYSCALL_DEFINE1(acct, const char __user *, name) 262 { 263 int error = 0; 264 265 if (!capable(CAP_SYS_PACCT)) 266 return -EPERM; 267 268 if (name) { 269 struct filename *tmp = getname(name); 270 if (IS_ERR(tmp)) 271 return PTR_ERR(tmp); 272 mutex_lock(&acct_on_mutex); 273 error = acct_on(tmp); 274 mutex_unlock(&acct_on_mutex); 275 putname(tmp); 276 } else { 277 acct_kill(acct_get(task_active_pid_ns(current)), NULL); 278 } 279 280 return error; 281 } 282 283 void acct_auto_close_mnt(struct hlist_head *list) 284 { 285 while (1) { 286 spin_lock(&acct_lock); 287 if (!list->first) 288 break; 289 acct_kill(__acct_get(hlist_entry(list->first, 290 struct bsd_acct_struct, 291 m_list)), NULL); 292 } 293 spin_unlock(&acct_lock); 294 } 295 296 void acct_auto_close(struct hlist_head *list) 297 { 298 while (1) { 299 spin_lock(&acct_lock); 300 if (!list->first) 301 break; 302 acct_kill(__acct_get(hlist_entry(list->first, 303 struct bsd_acct_struct, 304 s_list)), NULL); 305 } 306 spin_unlock(&acct_lock); 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) 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 ac->ac_etime_hi = etime >> 16; 456 ac->ac_etime_lo = (u16) etime; 457 } 458 #endif 459 do_div(elapsed, AHZ); 460 ac->ac_btime = get_seconds() - elapsed; 461 #if ACCT_VERSION==2 462 ac->ac_ahz = AHZ; 463 #endif 464 465 spin_lock_irq(¤t->sighand->siglock); 466 tty = current->signal->tty; /* Safe as we hold the siglock */ 467 ac->ac_tty = tty ? old_encode_dev(tty_devnum(tty)) : 0; 468 ac->ac_utime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_utime))); 469 ac->ac_stime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_stime))); 470 ac->ac_flag = pacct->ac_flag; 471 ac->ac_mem = encode_comp_t(pacct->ac_mem); 472 ac->ac_minflt = encode_comp_t(pacct->ac_minflt); 473 ac->ac_majflt = encode_comp_t(pacct->ac_majflt); 474 ac->ac_exitcode = pacct->ac_exitcode; 475 spin_unlock_irq(¤t->sighand->siglock); 476 } 477 /* 478 * do_acct_process does all actual work. Caller holds the reference to file. 479 */ 480 static void do_acct_process(struct bsd_acct_struct *acct) 481 { 482 acct_t ac; 483 unsigned long flim; 484 const struct cred *orig_cred; 485 struct pid_namespace *ns = acct->ns; 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 ac.ac_pid = task_tgid_nr_ns(current, ns); 514 rcu_read_lock(); 515 ac.ac_ppid = task_tgid_nr_ns(rcu_dereference(current->real_parent), ns); 516 rcu_read_unlock(); 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 down_read(¤t->mm->mmap_sem); 547 vma = current->mm->mmap; 548 while (vma) { 549 vsize += vma->vm_end - vma->vm_start; 550 vma = vma->vm_next; 551 } 552 up_read(¤t->mm->mmap_sem); 553 } 554 555 spin_lock_irq(¤t->sighand->siglock); 556 if (group_dead) 557 pacct->ac_mem = vsize / 1024; 558 if (thread_group_leader(current)) { 559 pacct->ac_exitcode = exitcode; 560 if (current->flags & PF_FORKNOEXEC) 561 pacct->ac_flag |= AFORK; 562 } 563 if (current->flags & PF_SUPERPRIV) 564 pacct->ac_flag |= ASU; 565 if (current->flags & PF_DUMPCORE) 566 pacct->ac_flag |= ACORE; 567 if (current->flags & PF_SIGNALED) 568 pacct->ac_flag |= AXSIG; 569 task_cputime(current, &utime, &stime); 570 pacct->ac_utime += utime; 571 pacct->ac_stime += stime; 572 pacct->ac_minflt += current->min_flt; 573 pacct->ac_majflt += current->maj_flt; 574 spin_unlock_irq(¤t->sighand->siglock); 575 } 576 577 static void slow_acct_process(struct pid_namespace *ns) 578 { 579 for ( ; ns; ns = ns->parent) { 580 struct bsd_acct_struct *acct = acct_get(ns); 581 if (acct) { 582 do_acct_process(acct); 583 mutex_unlock(&acct->lock); 584 acct_put(acct); 585 } 586 } 587 } 588 589 /** 590 * acct_process 591 * 592 * handles process accounting for an exiting task 593 */ 594 void acct_process(void) 595 { 596 struct pid_namespace *ns; 597 598 /* 599 * This loop is safe lockless, since current is still 600 * alive and holds its namespace, which in turn holds 601 * its parent. 602 */ 603 for (ns = task_active_pid_ns(current); ns != NULL; ns = ns->parent) { 604 if (ns->bacct) 605 break; 606 } 607 if (unlikely(ns)) 608 slow_acct_process(ns); 609 } 610