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 <linux/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 bsd_acct_struct *acct,
  79                 struct pid_namespace *ns, struct file *);
  80
  81 /*
  82  * This structure is used so that all the data protected by lock
  83  * can be placed in the same cache line as the lock.  This primes
  84  * the cache line to have the data after getting the lock.
  85  */
  86 struct bsd_acct_struct {
  87         int                     active;
  88         unsigned long           needcheck;
  89         struct file             *file;
  90         struct pid_namespace    *ns;
  91         struct list_head        list;
  92 };
  93
  94 static DEFINE_SPINLOCK(acct_lock);
  95 static LIST_HEAD(acct_list);
  96
  97 /*
  98  * Check the amount of free space and suspend/resume accordingly.
  99  */
 100 static int check_free_space(struct bsd_acct_struct *acct, struct file *file)
 101 {
 102         struct kstatfs sbuf;
 103         int res;
 104         int act;
 105         u64 resume;
 106         u64 suspend;
 107
 108         spin_lock(&acct_lock);
 109         res = acct->active;
 110         if (!file || time_is_before_jiffies(acct->needcheck))
 111                 goto out;
 112         spin_unlock(&acct_lock);
 113
 114         /* May block */
 115         if (vfs_statfs(&file->f_path, &sbuf))
 116                 return res;
 117         suspend = sbuf.f_blocks * SUSPEND;
 118         resume = sbuf.f_blocks * RESUME;
 119
 120         do_div(suspend, 100);
 121         do_div(resume, 100);
 122
 123         if (sbuf.f_bavail <= suspend)
 124                 act = -1;
 125         else if (sbuf.f_bavail >= resume)
 126                 act = 1;
 127         else
 128                 act = 0;
 129
 130         /*
 131          * If some joker switched acct->file under us we'ld better be
 132          * silent and _not_ touch anything.
 133          */
 134         spin_lock(&acct_lock);
 135         if (file != acct->file) {
 136                 if (act)
 137                         res = act > 0;
 138                 goto out;
 139         }
 140
 141         if (acct->active) {
 142                 if (act < 0) {
 143                         acct->active = 0;
 144                         printk(KERN_INFO "Process accounting paused\n");
 145                 }
 146         } else {
 147                 if (act > 0) {
 148                         acct->active = 1;
 149                         printk(KERN_INFO "Process accounting resumed\n");
 150                 }
 151         }
 152
 153         acct->needcheck = jiffies + ACCT_TIMEOUT*HZ;
 154         res = acct->active;
 155 out:
 156         spin_unlock(&acct_lock);
 157         return res;
 158 }
 159
 160 /*
 161  * Close the old accounting file (if currently open) and then replace
 162  * it with file (if non-NULL).
 163  *
 164  * NOTE: acct_lock MUST be held on entry and exit.
 165  */
 166 static void acct_file_reopen(struct bsd_acct_struct *acct, struct file *file,
 167                 struct pid_namespace *ns)
 168 {
 169         struct file *old_acct = NULL;
 170         struct pid_namespace *old_ns = NULL;
 171
 172         if (acct->file) {
 173                 old_acct = acct->file;
 174                 old_ns = acct->ns;
 175                 acct->active = 0;
 176                 acct->file = NULL;
 177                 acct->ns = NULL;
 178                 list_del(&acct->list);
 179         }
 180         if (file) {
 181                 acct->file = file;
 182                 acct->ns = ns;
 183                 acct->needcheck = jiffies + ACCT_TIMEOUT*HZ;
 184                 acct->active = 1;
 185                 list_add(&acct->list, &acct_list);
 186         }
 187         if (old_acct) {
 188                 mnt_unpin(old_acct->f_path.mnt);
 189                 spin_unlock(&acct_lock);
 190                 do_acct_process(acct, old_ns, old_acct);
 191                 filp_close(old_acct, NULL);
 192                 spin_lock(&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;
 201         struct bsd_acct_struct *acct = NULL;
 202
 203         /* Difference from BSD - they don't do O_APPEND */
 204         file = file_open_name(pathname, O_WRONLY|O_APPEND|O_LARGEFILE, 0);
 205         if (IS_ERR(file))
 206                 return PTR_ERR(file);
 207
 208         if (!S_ISREG(file_inode(file)->i_mode)) {
 209                 filp_close(file, NULL);
 210                 return -EACCES;
 211         }
 212
 213         if (!file->f_op->write) {
 214                 filp_close(file, NULL);
 215                 return -EIO;
 216         }
 217
 218         ns = task_active_pid_ns(current);
 219         if (ns->bacct == NULL) {
 220                 acct = kzalloc(sizeof(struct bsd_acct_struct), GFP_KERNEL);
 221                 if (acct == NULL) {
 222                         filp_close(file, NULL);
 223                         return -ENOMEM;
 224                 }
 225         }
 226
 227         spin_lock(&acct_lock);
 228         if (ns->bacct == NULL) {
 229                 ns->bacct = acct;
 230                 acct = NULL;
 231         }
 232
 233         mnt = file->f_path.mnt;
 234         mnt_pin(mnt);
 235         acct_file_reopen(ns->bacct, file, ns);
 236         spin_unlock(&acct_lock);
 237
 238         mntput(mnt); /* it's pinned, now give up active reference */
 239         kfree(acct);
 240
 241         return 0;
 242 }
 243
 244 /**
 245  * sys_acct - enable/disable process accounting
 246  * @name: file name for accounting records or NULL to shutdown accounting
 247  *
 248  * Returns 0 for success or negative errno values for failure.
 249  *
 250  * sys_acct() is the only system call needed to implement process
 251  * accounting. It takes the name of the file where accounting records
 252  * should be written. If the filename is NULL, accounting will be
 253  * shutdown.
 254  */
 255 SYSCALL_DEFINE1(acct, const char __user *, name)
 256 {
 257         int error = 0;
 258
 259         if (!capable(CAP_SYS_PACCT))
 260                 return -EPERM;
 261
 262         if (name) {
 263                 struct filename *tmp = getname(name);
 264                 if (IS_ERR(tmp))
 265                         return PTR_ERR(tmp);
 266                 error = acct_on(tmp);
 267                 putname(tmp);
 268         } else {
 269                 struct bsd_acct_struct *acct;
 270
 271                 acct = task_active_pid_ns(current)->bacct;
 272                 if (acct == NULL)
 273                         return 0;
 274
 275                 spin_lock(&acct_lock);
 276                 acct_file_reopen(acct, NULL, NULL);
 277                 spin_unlock(&acct_lock);
 278         }
 279
 280         return error;
 281 }
 282
 283 /**
 284  * acct_auto_close - turn off a filesystem's accounting if it is on
 285  * @m: vfsmount being shut down
 286  *
 287  * If the accounting is turned on for a file in the subtree pointed to
 288  * to by m, turn accounting off.  Done when m is about to die.
 289  */
 290 void acct_auto_close_mnt(struct vfsmount *m)
 291 {
 292         struct bsd_acct_struct *acct;
 293
 294         spin_lock(&acct_lock);
 295 restart:
 296         list_for_each_entry(acct, &acct_list, list)
 297                 if (acct->file && acct->file->f_path.mnt == m) {
 298                         acct_file_reopen(acct, NULL, NULL);
 299                         goto restart;
 300                 }
 301         spin_unlock(&acct_lock);
 302 }
 303
 304 /**
 305  * acct_auto_close - turn off a filesystem's accounting if it is on
 306  * @sb: super block for the filesystem
 307  *
 308  * If the accounting is turned on for a file in the filesystem pointed
 309  * to by sb, turn accounting off.
 310  */
 311 void acct_auto_close(struct super_block *sb)
 312 {
 313         struct bsd_acct_struct *acct;
 314
 315         spin_lock(&acct_lock);
 316 restart:
 317         list_for_each_entry(acct, &acct_list, list)
 318                 if (acct->file && acct->file->f_path.dentry->d_sb == sb) {
 319                         acct_file_reopen(acct, NULL, NULL);
 320                         goto restart;
 321                 }
 322         spin_unlock(&acct_lock);
 323 }
 324
 325 void acct_exit_ns(struct pid_namespace *ns)
 326 {
 327         struct bsd_acct_struct *acct = ns->bacct;
 328
 329         if (acct == NULL)
 330                 return;
 331
 332         spin_lock(&acct_lock);
 333         if (acct->file != NULL)
 334                 acct_file_reopen(acct, NULL, NULL);
 335         spin_unlock(&acct_lock);
 336
 337         kfree(acct);
 338 }
 339
 340 /*
 341  *  encode an unsigned long into a comp_t
 342  *
 343  *  This routine has been adopted from the encode_comp_t() function in
 344  *  the kern_acct.c file of the FreeBSD operating system. The encoding
 345  *  is a 13-bit fraction with a 3-bit (base 8) exponent.
 346  */
 347
 348 #define MANTSIZE        13                      /* 13 bit mantissa. */
 349 #define EXPSIZE         3                       /* Base 8 (3 bit) exponent. */
 350 #define MAXFRACT        ((1 << MANTSIZE) - 1)   /* Maximum fractional value. */
 351
 352 static comp_t encode_comp_t(unsigned long value)
 353 {
 354         int exp, rnd;
 355
 356         exp = rnd = 0;
 357         while (value > MAXFRACT) {
 358                 rnd = value & (1 << (EXPSIZE - 1));     /* Round up? */
 359                 value >>= EXPSIZE;      /* Base 8 exponent == 3 bit shift. */
 360                 exp++;
 361         }
 362
 363         /*
 364          * If we need to round up, do it (and handle overflow correctly).
 365          */
 366         if (rnd && (++value > MAXFRACT)) {
 367                 value >>= EXPSIZE;
 368                 exp++;
 369         }
 370
 371         /*
 372          * Clean it up and polish it off.
 373          */
 374         exp <<= MANTSIZE;               /* Shift the exponent into place */
 375         exp += value;                   /* and add on the mantissa. */
 376         return exp;
 377 }
 378
 379 #if ACCT_VERSION==1 || ACCT_VERSION==2
 380 /*
 381  * encode an u64 into a comp2_t (24 bits)
 382  *
 383  * Format: 5 bit base 2 exponent, 20 bits mantissa.
 384  * The leading bit of the mantissa is not stored, but implied for
 385  * non-zero exponents.
 386  * Largest encodable value is 50 bits.
 387  */
 388
 389 #define MANTSIZE2       20                      /* 20 bit mantissa. */
 390 #define EXPSIZE2        5                       /* 5 bit base 2 exponent. */
 391 #define MAXFRACT2       ((1ul << MANTSIZE2) - 1) /* Maximum fractional value. */
 392 #define MAXEXP2         ((1 <<EXPSIZE2) - 1)    /* Maximum exponent. */
 393
 394 static comp2_t encode_comp2_t(u64 value)
 395 {
 396         int exp, rnd;
 397
 398         exp = (value > (MAXFRACT2>>1));
 399         rnd = 0;
 400         while (value > MAXFRACT2) {
 401                 rnd = value & 1;
 402                 value >>= 1;
 403                 exp++;
 404         }
 405
 406         /*
 407          * If we need to round up, do it (and handle overflow correctly).
 408          */
 409         if (rnd && (++value > MAXFRACT2)) {
 410                 value >>= 1;
 411                 exp++;
 412         }
 413
 414         if (exp > MAXEXP2) {
 415                 /* Overflow. Return largest representable number instead. */
 416                 return (1ul << (MANTSIZE2+EXPSIZE2-1)) - 1;
 417         } else {
 418                 return (value & (MAXFRACT2>>1)) | (exp << (MANTSIZE2-1));
 419         }
 420 }
 421 #endif
 422
 423 #if ACCT_VERSION==3
 424 /*
 425  * encode an u64 into a 32 bit IEEE float
 426  */
 427 static u32 encode_float(u64 value)
 428 {
 429         unsigned exp = 190;
 430         unsigned u;
 431
 432         if (value==0) return 0;
 433         while ((s64)value > 0){
 434                 value <<= 1;
 435                 exp--;
 436         }
 437         u = (u32)(value >> 40) & 0x7fffffu;
 438         return u | (exp << 23);
 439 }
 440 #endif
 441
 442 /*
 443  *  Write an accounting entry for an exiting process
 444  *
 445  *  The acct_process() call is the workhorse of the process
 446  *  accounting system. The struct acct is built here and then written
 447  *  into the accounting file. This function should only be called from
 448  *  do_exit() or when switching to a different output file.
 449  */
 450
 451 /*
 452  *  do_acct_process does all actual work. Caller holds the reference to file.
 453  */
 454 static void do_acct_process(struct bsd_acct_struct *acct,
 455                 struct pid_namespace *ns, struct file *file)
 456 {
 457         struct pacct_struct *pacct = &current->signal->pacct;
 458         acct_t ac;
 459         mm_segment_t fs;
 460         unsigned long flim;
 461         u64 elapsed;
 462         u64 run_time;
 463         struct timespec uptime;
 464         struct tty_struct *tty;
 465         const struct cred *orig_cred;
 466
 467         /* Perform file operations on behalf of whoever enabled accounting */
 468         orig_cred = override_creds(file->f_cred);
 469
 470         /*
 471          * First check to see if there is enough free_space to continue
 472          * the process accounting system.
 473          */
 474         if (!check_free_space(acct, file))
 475                 goto out;
 476
 477         /*
 478          * Fill the accounting struct with the needed info as recorded
 479          * by the different kernel functions.
 480          */
 481         memset(&ac, 0, sizeof(acct_t));
 482
 483         ac.ac_version = ACCT_VERSION | ACCT_BYTEORDER;
 484         strlcpy(ac.ac_comm, current->comm, sizeof(ac.ac_comm));
 485
 486         /* calculate run_time in nsec*/
 487         do_posix_clock_monotonic_gettime(&uptime);
 488         run_time = (u64)uptime.tv_sec*NSEC_PER_SEC + uptime.tv_nsec;
 489         run_time -= (u64)current->group_leader->start_time.tv_sec * NSEC_PER_SEC
 490                        + current->group_leader->start_time.tv_nsec;
 491         /* convert nsec -> AHZ */
 492         elapsed = nsec_to_AHZ(run_time);
 493 #if ACCT_VERSION==3
 494         ac.ac_etime = encode_float(elapsed);
 495 #else
 496         ac.ac_etime = encode_comp_t(elapsed < (unsigned long) -1l ?
 497                                (unsigned long) elapsed : (unsigned long) -1l);
 498 #endif
 499 #if ACCT_VERSION==1 || ACCT_VERSION==2
 500         {
 501                 /* new enlarged etime field */
 502                 comp2_t etime = encode_comp2_t(elapsed);
 503                 ac.ac_etime_hi = etime >> 16;
 504                 ac.ac_etime_lo = (u16) etime;
 505         }
 506 #endif
 507         do_div(elapsed, AHZ);
 508         ac.ac_btime = get_seconds() - elapsed;
 509         /* we really need to bite the bullet and change layout */
 510         ac.ac_uid = from_kuid_munged(file->f_cred->user_ns, orig_cred->uid);
 511         ac.ac_gid = from_kgid_munged(file->f_cred->user_ns, orig_cred->gid);
 512 #if ACCT_VERSION==2
 513         ac.ac_ahz = AHZ;
 514 #endif
 515 #if ACCT_VERSION==1 || ACCT_VERSION==2
 516         /* backward-compatible 16 bit fields */
 517         ac.ac_uid16 = ac.ac_uid;
 518         ac.ac_gid16 = ac.ac_gid;
 519 #endif
 520 #if ACCT_VERSION==3
 521         ac.ac_pid = task_tgid_nr_ns(current, ns);
 522         rcu_read_lock();
 523         ac.ac_ppid = task_tgid_nr_ns(rcu_dereference(current->real_parent), ns);
 524         rcu_read_unlock();
 525 #endif
 526
 527         spin_lock_irq(&current->sighand->siglock);
 528         tty = current->signal->tty;     /* Safe as we hold the siglock */
 529         ac.ac_tty = tty ? old_encode_dev(tty_devnum(tty)) : 0;
 530         ac.ac_utime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_utime)));
 531         ac.ac_stime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_stime)));
 532         ac.ac_flag = pacct->ac_flag;
 533         ac.ac_mem = encode_comp_t(pacct->ac_mem);
 534         ac.ac_minflt = encode_comp_t(pacct->ac_minflt);
 535         ac.ac_majflt = encode_comp_t(pacct->ac_majflt);
 536         ac.ac_exitcode = pacct->ac_exitcode;
 537         spin_unlock_irq(&current->sighand->siglock);
 538         ac.ac_io = encode_comp_t(0 /* current->io_usage */);    /* %% */
 539         ac.ac_rw = encode_comp_t(ac.ac_io / 1024);
 540         ac.ac_swaps = encode_comp_t(0);
 541
 542         /*
 543          * Get freeze protection. If the fs is frozen, just skip the write
 544          * as we could deadlock the system otherwise.
 545          */
 546         if (!file_start_write_trylock(file))
 547                 goto out;
 548         /*
 549          * Kernel segment override to datasegment and write it
 550          * to the accounting file.
 551          */
 552         fs = get_fs();
 553         set_fs(KERNEL_DS);
 554         /*
 555          * Accounting records are not subject to resource limits.
 556          */
 557         flim = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
 558         current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
 559         file->f_op->write(file, (char *)&ac,
 560                                sizeof(acct_t), &file->f_pos);
 561         current->signal->rlim[RLIMIT_FSIZE].rlim_cur = flim;
 562         set_fs(fs);
 563         file_end_write(file);
 564 out:
 565         revert_creds(orig_cred);
 566 }
 567
 568 /**
 569  * acct_collect - collect accounting information into pacct_struct
 570  * @exitcode: task exit code
 571  * @group_dead: not 0, if this thread is the last one in the process.
 572  */
 573 void acct_collect(long exitcode, int group_dead)
 574 {
 575         struct pacct_struct *pacct = &current->signal->pacct;
 576         cputime_t utime, stime;
 577         unsigned long vsize = 0;
 578
 579         if (group_dead && current->mm) {
 580                 struct vm_area_struct *vma;
 581                 down_read(&current->mm->mmap_sem);
 582                 vma = current->mm->mmap;
 583                 while (vma) {
 584                         vsize += vma->vm_end - vma->vm_start;
 585                         vma = vma->vm_next;
 586                 }
 587                 up_read(&current->mm->mmap_sem);
 588         }
 589
 590         spin_lock_irq(&current->sighand->siglock);
 591         if (group_dead)
 592                 pacct->ac_mem = vsize / 1024;
 593         if (thread_group_leader(current)) {
 594                 pacct->ac_exitcode = exitcode;
 595                 if (current->flags & PF_FORKNOEXEC)
 596                         pacct->ac_flag |= AFORK;
 597         }
 598         if (current->flags & PF_SUPERPRIV)
 599                 pacct->ac_flag |= ASU;
 600         if (current->flags & PF_DUMPCORE)
 601                 pacct->ac_flag |= ACORE;
 602         if (current->flags & PF_SIGNALED)
 603                 pacct->ac_flag |= AXSIG;
 604         task_cputime(current, &utime, &stime);
 605         pacct->ac_utime += utime;
 606         pacct->ac_stime += stime;
 607         pacct->ac_minflt += current->min_flt;
 608         pacct->ac_majflt += current->maj_flt;
 609         spin_unlock_irq(&current->sighand->siglock);
 610 }
 611
 612 static void acct_process_in_ns(struct pid_namespace *ns)
 613 {
 614         struct file *file = NULL;
 615         struct bsd_acct_struct *acct;
 616
 617         acct = ns->bacct;
 618         /*
 619          * accelerate the common fastpath:
 620          */
 621         if (!acct || !acct->file)
 622                 return;
 623
 624         spin_lock(&acct_lock);
 625         file = acct->file;
 626         if (unlikely(!file)) {
 627                 spin_unlock(&acct_lock);
 628                 return;
 629         }
 630         get_file(file);
 631         spin_unlock(&acct_lock);
 632
 633         do_acct_process(acct, ns, file);
 634         fput(file);
 635 }
 636
 637 /**
 638  * acct_process - now just a wrapper around acct_process_in_ns,
 639  * which in turn is a wrapper around do_acct_process.
 640  *
 641  * handles process accounting for an exiting task
 642  */
 643 void acct_process(void)
 644 {
 645         struct pid_namespace *ns;
 646
 647         /*
 648          * This loop is safe lockless, since current is still
 649          * alive and holds its namespace, which in turn holds
 650          * its parent.
 651          */
 652         for (ns = task_active_pid_ns(current); ns != NULL; ns = ns->parent)
 653                 acct_process_in_ns(ns);
 654 }