kernel/acct.c

   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 <asm/uaccess.h>
  59 #include <asm/div64.h>
  60 #include <linux/blkdev.h> /* sector_div */
  61 #include <linux/pid_namespace.h>
  62
  63 /*
  64  * These constants control the amount of freespace that suspend and
  65  * resume the process accounting system, and the time delay between
  66  * each check.
  67  * Turned into sysctl-controllable parameters. AV, 12/11/98
  68  */
  69
  70 int acct_parm[3] = {4, 2, 30};
  71 #define RESUME          (acct_parm[0])  /* >foo% free space - resume */
  72 #define SUSPEND         (acct_parm[1])  /* <foo% free space - suspend */
  73 #define ACCT_TIMEOUT    (acct_parm[2])  /* foo second timeout between checks */
  74
  75 /*
  76  * External references and all of the globals.
  77  */
  78 static void do_acct_process(struct pid_namespace *ns, struct file *);
  79
  80 /*
  81  * This structure is used so that all the data protected by lock
  82  * can be placed in the same cache line as the lock.  This primes
  83  * the cache line to have the data after getting the lock.
  84  */
  85 struct acct_glbs {
  86         spinlock_t              lock;
  87         volatile int            active;
  88         volatile int            needcheck;
  89         struct file             *file;
  90         struct pid_namespace    *ns;
  91         struct timer_list       timer;
  92 };
  93
  94 static struct acct_glbs acct_globals __cacheline_aligned =
  95         {__SPIN_LOCK_UNLOCKED(acct_globals.lock)};
  96
  97 /*
  98  * Called whenever the timer says to check the free space.
  99  */
 100 static void acct_timeout(unsigned long unused)
 101 {
 102         acct_globals.needcheck = 1;
 103 }
 104
 105 /*
 106  * Check the amount of free space and suspend/resume accordingly.
 107  */
 108 static int check_free_space(struct file *file)
 109 {
 110         struct kstatfs sbuf;
 111         int res;
 112         int act;
 113         sector_t resume;
 114         sector_t suspend;
 115
 116         spin_lock(&acct_globals.lock);
 117         res = acct_globals.active;
 118         if (!file || !acct_globals.needcheck)
 119                 goto out;
 120         spin_unlock(&acct_globals.lock);
 121
 122         /* May block */
 123         if (vfs_statfs(file->f_path.dentry, &sbuf))
 124                 return res;
 125         suspend = sbuf.f_blocks * SUSPEND;
 126         resume = sbuf.f_blocks * RESUME;
 127
 128         sector_div(suspend, 100);
 129         sector_div(resume, 100);
 130
 131         if (sbuf.f_bavail <= suspend)
 132                 act = -1;
 133         else if (sbuf.f_bavail >= resume)
 134                 act = 1;
 135         else
 136                 act = 0;
 137
 138         /*
 139          * If some joker switched acct_globals.file under us we'ld better be
 140          * silent and _not_ touch anything.
 141          */
 142         spin_lock(&acct_globals.lock);
 143         if (file != acct_globals.file) {
 144                 if (act)
 145                         res = act>0;
 146                 goto out;
 147         }
 148
 149         if (acct_globals.active) {
 150                 if (act < 0) {
 151                         acct_globals.active = 0;
 152                         printk(KERN_INFO "Process accounting paused\n");
 153                 }
 154         } else {
 155                 if (act > 0) {
 156                         acct_globals.active = 1;
 157                         printk(KERN_INFO "Process accounting resumed\n");
 158                 }
 159         }
 160
 161         del_timer(&acct_globals.timer);
 162         acct_globals.needcheck = 0;
 163         acct_globals.timer.expires = jiffies + ACCT_TIMEOUT*HZ;
 164         add_timer(&acct_globals.timer);
 165         res = acct_globals.active;
 166 out:
 167         spin_unlock(&acct_globals.lock);
 168         return res;
 169 }
 170
 171 /*
 172  * Close the old accounting file (if currently open) and then replace
 173  * it with file (if non-NULL).
 174  *
 175  * NOTE: acct_globals.lock MUST be held on entry and exit.
 176  */
 177 static void acct_file_reopen(struct file *file)
 178 {
 179         struct file *old_acct = NULL;
 180         struct pid_namespace *old_ns = NULL;
 181
 182         if (acct_globals.file) {
 183                 old_acct = acct_globals.file;
 184                 old_ns = acct_globals.ns;
 185                 del_timer(&acct_globals.timer);
 186                 acct_globals.active = 0;
 187                 acct_globals.needcheck = 0;
 188                 acct_globals.file = NULL;
 189         }
 190         if (file) {
 191                 acct_globals.file = file;
 192                 acct_globals.ns = get_pid_ns(task_active_pid_ns(current));
 193                 acct_globals.needcheck = 0;
 194                 acct_globals.active = 1;
 195                 /* It's been deleted if it was used before so this is safe */
 196                 init_timer(&acct_globals.timer);
 197                 acct_globals.timer.function = acct_timeout;
 198                 acct_globals.timer.expires = jiffies + ACCT_TIMEOUT*HZ;
 199                 add_timer(&acct_globals.timer);
 200         }
 201         if (old_acct) {
 202                 mnt_unpin(old_acct->f_path.mnt);
 203                 spin_unlock(&acct_globals.lock);
 204                 do_acct_process(old_ns, old_acct);
 205                 filp_close(old_acct, NULL);
 206                 put_pid_ns(old_ns);
 207                 spin_lock(&acct_globals.lock);
 208         }
 209 }
 210
 211 static int acct_on(char *name)
 212 {
 213         struct file *file;
 214         int error;
 215
 216         /* Difference from BSD - they don't do O_APPEND */
 217         file = filp_open(name, O_WRONLY|O_APPEND|O_LARGEFILE, 0);
 218         if (IS_ERR(file))
 219                 return PTR_ERR(file);
 220
 221         if (!S_ISREG(file->f_path.dentry->d_inode->i_mode)) {
 222                 filp_close(file, NULL);
 223                 return -EACCES;
 224         }
 225
 226         if (!file->f_op->write) {
 227                 filp_close(file, NULL);
 228                 return -EIO;
 229         }
 230
 231         error = security_acct(file);
 232         if (error) {
 233                 filp_close(file, NULL);
 234                 return error;
 235         }
 236
 237         spin_lock(&acct_globals.lock);
 238         mnt_pin(file->f_path.mnt);
 239         acct_file_reopen(file);
 240         spin_unlock(&acct_globals.lock);
 241
 242         mntput(file->f_path.mnt); /* it's pinned, now give up active reference */
 243
 244         return 0;
 245 }
 246
 247 /**
 248  * sys_acct - enable/disable process accounting
 249  * @name: file name for accounting records or NULL to shutdown accounting
 250  *
 251  * Returns 0 for success or negative errno values for failure.
 252  *
 253  * sys_acct() is the only system call needed to implement process
 254  * accounting. It takes the name of the file where accounting records
 255  * should be written. If the filename is NULL, accounting will be
 256  * shutdown.
 257  */
 258 asmlinkage long sys_acct(const char __user *name)
 259 {
 260         int error;
 261
 262         if (!capable(CAP_SYS_PACCT))
 263                 return -EPERM;
 264
 265         if (name) {
 266                 char *tmp = getname(name);
 267                 if (IS_ERR(tmp))
 268                         return (PTR_ERR(tmp));
 269                 error = acct_on(tmp);
 270                 putname(tmp);
 271         } else {
 272                 error = security_acct(NULL);
 273                 if (!error) {
 274                         spin_lock(&acct_globals.lock);
 275                         acct_file_reopen(NULL);
 276                         spin_unlock(&acct_globals.lock);
 277                 }
 278         }
 279         return error;
 280 }
 281
 282 /**
 283  * acct_auto_close - turn off a filesystem's accounting if it is on
 284  * @m: vfsmount being shut down
 285  *
 286  * If the accounting is turned on for a file in the subtree pointed to
 287  * to by m, turn accounting off.  Done when m is about to die.
 288  */
 289 void acct_auto_close_mnt(struct vfsmount *m)
 290 {
 291         spin_lock(&acct_globals.lock);
 292         if (acct_globals.file && acct_globals.file->f_path.mnt == m)
 293                 acct_file_reopen(NULL);
 294         spin_unlock(&acct_globals.lock);
 295 }
 296
 297 /**
 298  * acct_auto_close - turn off a filesystem's accounting if it is on
 299  * @sb: super block for the filesystem
 300  *
 301  * If the accounting is turned on for a file in the filesystem pointed
 302  * to by sb, turn accounting off.
 303  */
 304 void acct_auto_close(struct super_block *sb)
 305 {
 306         spin_lock(&acct_globals.lock);
 307         if (acct_globals.file &&
 308             acct_globals.file->f_path.mnt->mnt_sb == sb) {
 309                 acct_file_reopen(NULL);
 310         }
 311         spin_unlock(&acct_globals.lock);
 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 /*
 426  *  do_acct_process does all actual work. Caller holds the reference to file.
 427  */
 428 static void do_acct_process(struct pid_namespace *ns, struct file *file)
 429 {
 430         struct pacct_struct *pacct = &current->signal->pacct;
 431         acct_t ac;
 432         mm_segment_t fs;
 433         unsigned long flim;
 434         u64 elapsed;
 435         u64 run_time;
 436         struct timespec uptime;
 437         struct tty_struct *tty;
 438
 439         /*
 440          * First check to see if there is enough free_space to continue
 441          * the process accounting system.
 442          */
 443         if (!check_free_space(file))
 444                 return;
 445
 446         /*
 447          * Fill the accounting struct with the needed info as recorded
 448          * by the different kernel functions.
 449          */
 450         memset((caddr_t)&ac, 0, sizeof(acct_t));
 451
 452         ac.ac_version = ACCT_VERSION | ACCT_BYTEORDER;
 453         strlcpy(ac.ac_comm, current->comm, sizeof(ac.ac_comm));
 454
 455         /* calculate run_time in nsec*/
 456         do_posix_clock_monotonic_gettime(&uptime);
 457         run_time = (u64)uptime.tv_sec*NSEC_PER_SEC + uptime.tv_nsec;
 458         run_time -= (u64)current->group_leader->start_time.tv_sec * NSEC_PER_SEC
 459                        + current->group_leader->start_time.tv_nsec;
 460         /* convert nsec -> AHZ */
 461         elapsed = nsec_to_AHZ(run_time);
 462 #if ACCT_VERSION==3
 463         ac.ac_etime = encode_float(elapsed);
 464 #else
 465         ac.ac_etime = encode_comp_t(elapsed < (unsigned long) -1l ?
 466                                (unsigned long) elapsed : (unsigned long) -1l);
 467 #endif
 468 #if ACCT_VERSION==1 || ACCT_VERSION==2
 469         {
 470                 /* new enlarged etime field */
 471                 comp2_t etime = encode_comp2_t(elapsed);
 472                 ac.ac_etime_hi = etime >> 16;
 473                 ac.ac_etime_lo = (u16) etime;
 474         }
 475 #endif
 476         do_div(elapsed, AHZ);
 477         ac.ac_btime = get_seconds() - elapsed;
 478         /* we really need to bite the bullet and change layout */
 479         ac.ac_uid = current->uid;
 480         ac.ac_gid = current->gid;
 481 #if ACCT_VERSION==2
 482         ac.ac_ahz = AHZ;
 483 #endif
 484 #if ACCT_VERSION==1 || ACCT_VERSION==2
 485         /* backward-compatible 16 bit fields */
 486         ac.ac_uid16 = current->uid;
 487         ac.ac_gid16 = current->gid;
 488 #endif
 489 #if ACCT_VERSION==3
 490         ac.ac_pid = task_tgid_nr_ns(current, ns);
 491         rcu_read_lock();
 492         ac.ac_ppid = task_tgid_nr_ns(rcu_dereference(current->real_parent), ns);
 493         rcu_read_unlock();
 494 #endif
 495
 496         spin_lock_irq(&current->sighand->siglock);
 497         tty = current->signal->tty;
 498         ac.ac_tty = tty ? old_encode_dev(tty_devnum(tty)) : 0;
 499         ac.ac_utime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_utime)));
 500         ac.ac_stime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_stime)));
 501         ac.ac_flag = pacct->ac_flag;
 502         ac.ac_mem = encode_comp_t(pacct->ac_mem);
 503         ac.ac_minflt = encode_comp_t(pacct->ac_minflt);
 504         ac.ac_majflt = encode_comp_t(pacct->ac_majflt);
 505         ac.ac_exitcode = pacct->ac_exitcode;
 506         spin_unlock_irq(&current->sighand->siglock);
 507         ac.ac_io = encode_comp_t(0 /* current->io_usage */);    /* %% */
 508         ac.ac_rw = encode_comp_t(ac.ac_io / 1024);
 509         ac.ac_swaps = encode_comp_t(0);
 510
 511         /*
 512          * Kernel segment override to datasegment and write it
 513          * to the accounting file.
 514          */
 515         fs = get_fs();
 516         set_fs(KERNEL_DS);
 517         /*
 518          * Accounting records are not subject to resource limits.
 519          */
 520         flim = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
 521         current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
 522         file->f_op->write(file, (char *)&ac,
 523                                sizeof(acct_t), &file->f_pos);
 524         current->signal->rlim[RLIMIT_FSIZE].rlim_cur = flim;
 525         set_fs(fs);
 526 }
 527
 528 /**
 529  * acct_init_pacct - initialize a new pacct_struct
 530  * @pacct: per-process accounting info struct to initialize
 531  */
 532 void acct_init_pacct(struct pacct_struct *pacct)
 533 {
 534         memset(pacct, 0, sizeof(struct pacct_struct));
 535         pacct->ac_utime = pacct->ac_stime = cputime_zero;
 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 = &current->signal->pacct;
 546         unsigned long vsize = 0;
 547
 548         if (group_dead && current->mm) {
 549                 struct vm_area_struct *vma;
 550                 down_read(&current->mm->mmap_sem);
 551                 vma = current->mm->mmap;
 552                 while (vma) {
 553                         vsize += vma->vm_end - vma->vm_start;
 554                         vma = vma->vm_next;
 555                 }
 556                 up_read(&current->mm->mmap_sem);
 557         }
 558
 559         spin_lock_irq(&current->sighand->siglock);
 560         if (group_dead)
 561                 pacct->ac_mem = vsize / 1024;
 562         if (thread_group_leader(current)) {
 563                 pacct->ac_exitcode = exitcode;
 564                 if (current->flags & PF_FORKNOEXEC)
 565                         pacct->ac_flag |= AFORK;
 566         }
 567         if (current->flags & PF_SUPERPRIV)
 568                 pacct->ac_flag |= ASU;
 569         if (current->flags & PF_DUMPCORE)
 570                 pacct->ac_flag |= ACORE;
 571         if (current->flags & PF_SIGNALED)
 572                 pacct->ac_flag |= AXSIG;
 573         pacct->ac_utime = cputime_add(pacct->ac_utime, current->utime);
 574         pacct->ac_stime = cputime_add(pacct->ac_stime, current->stime);
 575         pacct->ac_minflt += current->min_flt;
 576         pacct->ac_majflt += current->maj_flt;
 577         spin_unlock_irq(&current->sighand->siglock);
 578 }
 579
 580 /**
 581  * acct_process - now just a wrapper around do_acct_process
 582  * @exitcode: task exit code
 583  *
 584  * handles process accounting for an exiting task
 585  */
 586 void acct_process(void)
 587 {
 588         struct file *file = NULL;
 589         struct pid_namespace *ns;
 590
 591         /*
 592          * accelerate the common fastpath:
 593          */
 594         if (!acct_globals.file)
 595                 return;
 596
 597         spin_lock(&acct_globals.lock);
 598         file = acct_globals.file;
 599         if (unlikely(!file)) {
 600                 spin_unlock(&acct_globals.lock);
 601                 return;
 602         }
 603         get_file(file);
 604         ns = get_pid_ns(acct_globals.ns);
 605         spin_unlock(&acct_globals.lock);
 606
 607         do_acct_process(ns, file);
 608         fput(file);
 609         put_pid_ns(ns);
 610 }