4 * BSD Process Accounting for Linux
6 * Author: Marco van Wieringen <mvw@planets.elm.net>
8 * Some code based on ideas and code from:
9 * Thomas K. Dyas <tdyas@eden.rutgers.edu>
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.
17 * (C) Copyright 1995 - 1997 Marco van Wieringen - ELM Consultancy B.V.
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.
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.
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.
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.
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.
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>
64 * These constants control the amount of freespace that suspend and
65 * resume the process accounting system, and the time delay between
67 * Turned into sysctl-controllable parameters. AV, 12/11/98
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 */
76 * External references and all of the globals.
78 static void do_acct_process(struct bsd_acct_struct *acct);
80 struct bsd_acct_struct {
84 unsigned long needcheck;
86 struct pid_namespace *ns;
87 struct list_head list;
90 static DEFINE_SPINLOCK(acct_lock);
91 static LIST_HEAD(acct_list);
94 * Check the amount of free space and suspend/resume accordingly.
96 static int check_free_space(struct bsd_acct_struct *acct, struct file *file)
104 spin_lock(&acct_lock);
106 if (!file || time_is_before_jiffies(acct->needcheck))
108 spin_unlock(&acct_lock);
111 if (vfs_statfs(&file->f_path, &sbuf))
113 suspend = sbuf.f_blocks * SUSPEND;
114 resume = sbuf.f_blocks * RESUME;
116 do_div(suspend, 100);
119 if (sbuf.f_bavail <= suspend)
121 else if (sbuf.f_bavail >= resume)
127 * If some joker switched acct->file under us we'ld better be
128 * silent and _not_ touch anything.
130 spin_lock(&acct_lock);
131 if (file != acct->file) {
140 printk(KERN_INFO "Process accounting paused\n");
145 printk(KERN_INFO "Process accounting resumed\n");
149 acct->needcheck = jiffies + ACCT_TIMEOUT*HZ;
152 spin_unlock(&acct_lock);
156 static void acct_put(struct bsd_acct_struct *p)
158 spin_lock(&acct_lock);
161 spin_unlock(&acct_lock);
164 static struct bsd_acct_struct *acct_get(struct bsd_acct_struct **p)
166 struct bsd_acct_struct *res;
167 spin_lock(&acct_lock);
172 spin_unlock(&acct_lock);
174 mutex_lock(&res->lock);
176 mutex_unlock(&res->lock);
177 spin_lock(&acct_lock);
186 static void acct_kill(struct bsd_acct_struct *acct,
187 struct bsd_acct_struct *new)
190 struct file *file = acct->file;
191 struct pid_namespace *ns = acct->ns;
192 spin_lock(&acct_lock);
193 list_del(&acct->list);
194 mnt_unpin(file->f_path.mnt);
195 spin_unlock(&acct_lock);
196 do_acct_process(acct);
197 filp_close(file, NULL);
198 spin_lock(&acct_lock);
201 mnt_pin(new->file->f_path.mnt);
202 list_add(&new->list, &acct_list);
205 mutex_unlock(&acct->lock);
206 if (!(acct->count -= 2))
208 spin_unlock(&acct_lock);
212 static int acct_on(struct filename *pathname)
215 struct vfsmount *mnt;
216 struct pid_namespace *ns = task_active_pid_ns(current);
217 struct bsd_acct_struct *acct, *old;
219 acct = kzalloc(sizeof(struct bsd_acct_struct), GFP_KERNEL);
223 /* Difference from BSD - they don't do O_APPEND */
224 file = file_open_name(pathname, O_WRONLY|O_APPEND|O_LARGEFILE, 0);
227 return PTR_ERR(file);
230 if (!S_ISREG(file_inode(file)->i_mode)) {
232 filp_close(file, NULL);
236 if (!file->f_op->write) {
238 filp_close(file, NULL);
244 acct->needcheck = jiffies;
246 mutex_init(&acct->lock);
247 mnt = file->f_path.mnt;
249 old = acct_get(&ns->bacct);
251 acct_kill(old, acct);
253 spin_lock(&acct_lock);
256 list_add(&acct->list, &acct_list);
257 spin_unlock(&acct_lock);
259 mntput(mnt); /* it's pinned, now give up active reference */
263 static DEFINE_MUTEX(acct_on_mutex);
266 * sys_acct - enable/disable process accounting
267 * @name: file name for accounting records or NULL to shutdown accounting
269 * Returns 0 for success or negative errno values for failure.
271 * sys_acct() is the only system call needed to implement process
272 * accounting. It takes the name of the file where accounting records
273 * should be written. If the filename is NULL, accounting will be
276 SYSCALL_DEFINE1(acct, const char __user *, name)
280 if (!capable(CAP_SYS_PACCT))
284 struct filename *tmp = getname(name);
287 mutex_lock(&acct_on_mutex);
288 error = acct_on(tmp);
289 mutex_unlock(&acct_on_mutex);
292 acct_kill(acct_get(&task_active_pid_ns(current)->bacct), NULL);
299 * acct_auto_close - turn off a filesystem's accounting if it is on
300 * @m: vfsmount being shut down
302 * If the accounting is turned on for a file in the subtree pointed to
303 * to by m, turn accounting off. Done when m is about to die.
305 void acct_auto_close_mnt(struct vfsmount *m)
307 struct bsd_acct_struct *acct;
309 spin_lock(&acct_lock);
311 list_for_each_entry(acct, &acct_list, list)
312 if (acct->file->f_path.mnt == m) {
314 spin_unlock(&acct_lock);
315 mutex_lock(&acct->lock);
317 mutex_unlock(&acct->lock);
318 spin_lock(&acct_lock);
323 acct_kill(acct, NULL);
324 spin_lock(&acct_lock);
327 spin_unlock(&acct_lock);
331 * acct_auto_close - turn off a filesystem's accounting if it is on
332 * @sb: super block for the filesystem
334 * If the accounting is turned on for a file in the filesystem pointed
335 * to by sb, turn accounting off.
337 void acct_auto_close(struct super_block *sb)
339 struct bsd_acct_struct *acct;
341 spin_lock(&acct_lock);
343 list_for_each_entry(acct, &acct_list, list)
344 if (acct->file->f_path.dentry->d_sb == sb) {
346 spin_unlock(&acct_lock);
347 mutex_lock(&acct->lock);
349 mutex_unlock(&acct->lock);
350 spin_lock(&acct_lock);
355 acct_kill(acct, NULL);
356 spin_lock(&acct_lock);
359 spin_unlock(&acct_lock);
362 void acct_exit_ns(struct pid_namespace *ns)
364 acct_kill(acct_get(&ns->bacct), NULL);
368 * encode an unsigned long into a comp_t
370 * This routine has been adopted from the encode_comp_t() function in
371 * the kern_acct.c file of the FreeBSD operating system. The encoding
372 * is a 13-bit fraction with a 3-bit (base 8) exponent.
375 #define MANTSIZE 13 /* 13 bit mantissa. */
376 #define EXPSIZE 3 /* Base 8 (3 bit) exponent. */
377 #define MAXFRACT ((1 << MANTSIZE) - 1) /* Maximum fractional value. */
379 static comp_t encode_comp_t(unsigned long value)
384 while (value > MAXFRACT) {
385 rnd = value & (1 << (EXPSIZE - 1)); /* Round up? */
386 value >>= EXPSIZE; /* Base 8 exponent == 3 bit shift. */
391 * If we need to round up, do it (and handle overflow correctly).
393 if (rnd && (++value > MAXFRACT)) {
399 * Clean it up and polish it off.
401 exp <<= MANTSIZE; /* Shift the exponent into place */
402 exp += value; /* and add on the mantissa. */
406 #if ACCT_VERSION==1 || ACCT_VERSION==2
408 * encode an u64 into a comp2_t (24 bits)
410 * Format: 5 bit base 2 exponent, 20 bits mantissa.
411 * The leading bit of the mantissa is not stored, but implied for
412 * non-zero exponents.
413 * Largest encodable value is 50 bits.
416 #define MANTSIZE2 20 /* 20 bit mantissa. */
417 #define EXPSIZE2 5 /* 5 bit base 2 exponent. */
418 #define MAXFRACT2 ((1ul << MANTSIZE2) - 1) /* Maximum fractional value. */
419 #define MAXEXP2 ((1 <<EXPSIZE2) - 1) /* Maximum exponent. */
421 static comp2_t encode_comp2_t(u64 value)
425 exp = (value > (MAXFRACT2>>1));
427 while (value > MAXFRACT2) {
434 * If we need to round up, do it (and handle overflow correctly).
436 if (rnd && (++value > MAXFRACT2)) {
442 /* Overflow. Return largest representable number instead. */
443 return (1ul << (MANTSIZE2+EXPSIZE2-1)) - 1;
445 return (value & (MAXFRACT2>>1)) | (exp << (MANTSIZE2-1));
452 * encode an u64 into a 32 bit IEEE float
454 static u32 encode_float(u64 value)
459 if (value==0) return 0;
460 while ((s64)value > 0){
464 u = (u32)(value >> 40) & 0x7fffffu;
465 return u | (exp << 23);
470 * Write an accounting entry for an exiting process
472 * The acct_process() call is the workhorse of the process
473 * accounting system. The struct acct is built here and then written
474 * into the accounting file. This function should only be called from
475 * do_exit() or when switching to a different output file.
478 static void fill_ac(acct_t *ac)
480 struct pacct_struct *pacct = ¤t->signal->pacct;
481 u64 elapsed, run_time;
482 struct tty_struct *tty;
485 * Fill the accounting struct with the needed info as recorded
486 * by the different kernel functions.
488 memset(ac, 0, sizeof(acct_t));
490 ac->ac_version = ACCT_VERSION | ACCT_BYTEORDER;
491 strlcpy(ac->ac_comm, current->comm, sizeof(ac->ac_comm));
493 /* calculate run_time in nsec*/
494 run_time = ktime_get_ns();
495 run_time -= current->group_leader->start_time;
496 /* convert nsec -> AHZ */
497 elapsed = nsec_to_AHZ(run_time);
499 ac->ac_etime = encode_float(elapsed);
501 ac->ac_etime = encode_comp_t(elapsed < (unsigned long) -1l ?
502 (unsigned long) elapsed : (unsigned long) -1l);
504 #if ACCT_VERSION==1 || ACCT_VERSION==2
506 /* new enlarged etime field */
507 comp2_t etime = encode_comp2_t(elapsed);
508 ac->ac_etime_hi = etime >> 16;
509 ac->ac_etime_lo = (u16) etime;
512 do_div(elapsed, AHZ);
513 ac->ac_btime = get_seconds() - elapsed;
518 spin_lock_irq(¤t->sighand->siglock);
519 tty = current->signal->tty; /* Safe as we hold the siglock */
520 ac->ac_tty = tty ? old_encode_dev(tty_devnum(tty)) : 0;
521 ac->ac_utime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_utime)));
522 ac->ac_stime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_stime)));
523 ac->ac_flag = pacct->ac_flag;
524 ac->ac_mem = encode_comp_t(pacct->ac_mem);
525 ac->ac_minflt = encode_comp_t(pacct->ac_minflt);
526 ac->ac_majflt = encode_comp_t(pacct->ac_majflt);
527 ac->ac_exitcode = pacct->ac_exitcode;
528 spin_unlock_irq(¤t->sighand->siglock);
531 * do_acct_process does all actual work. Caller holds the reference to file.
533 static void do_acct_process(struct bsd_acct_struct *acct)
537 const struct cred *orig_cred;
538 struct pid_namespace *ns = acct->ns;
539 struct file *file = acct->file;
542 * Accounting records are not subject to resource limits.
544 flim = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
545 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
546 /* Perform file operations on behalf of whoever enabled accounting */
547 orig_cred = override_creds(file->f_cred);
550 * First check to see if there is enough free_space to continue
551 * the process accounting system.
553 if (!check_free_space(acct, file))
557 /* we really need to bite the bullet and change layout */
558 ac.ac_uid = from_kuid_munged(file->f_cred->user_ns, orig_cred->uid);
559 ac.ac_gid = from_kgid_munged(file->f_cred->user_ns, orig_cred->gid);
560 #if ACCT_VERSION==1 || ACCT_VERSION==2
561 /* backward-compatible 16 bit fields */
562 ac.ac_uid16 = ac.ac_uid;
563 ac.ac_gid16 = ac.ac_gid;
566 ac.ac_pid = task_tgid_nr_ns(current, ns);
568 ac.ac_ppid = task_tgid_nr_ns(rcu_dereference(current->real_parent), ns);
572 * Get freeze protection. If the fs is frozen, just skip the write
573 * as we could deadlock the system otherwise.
575 if (file_start_write_trylock(file)) {
576 /* it's been opened O_APPEND, so position is irrelevant */
578 __kernel_write(file, (char *)&ac, sizeof(acct_t), &pos);
579 file_end_write(file);
582 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = flim;
583 revert_creds(orig_cred);
587 * acct_collect - collect accounting information into pacct_struct
588 * @exitcode: task exit code
589 * @group_dead: not 0, if this thread is the last one in the process.
591 void acct_collect(long exitcode, int group_dead)
593 struct pacct_struct *pacct = ¤t->signal->pacct;
594 cputime_t utime, stime;
595 unsigned long vsize = 0;
597 if (group_dead && current->mm) {
598 struct vm_area_struct *vma;
599 down_read(¤t->mm->mmap_sem);
600 vma = current->mm->mmap;
602 vsize += vma->vm_end - vma->vm_start;
605 up_read(¤t->mm->mmap_sem);
608 spin_lock_irq(¤t->sighand->siglock);
610 pacct->ac_mem = vsize / 1024;
611 if (thread_group_leader(current)) {
612 pacct->ac_exitcode = exitcode;
613 if (current->flags & PF_FORKNOEXEC)
614 pacct->ac_flag |= AFORK;
616 if (current->flags & PF_SUPERPRIV)
617 pacct->ac_flag |= ASU;
618 if (current->flags & PF_DUMPCORE)
619 pacct->ac_flag |= ACORE;
620 if (current->flags & PF_SIGNALED)
621 pacct->ac_flag |= AXSIG;
622 task_cputime(current, &utime, &stime);
623 pacct->ac_utime += utime;
624 pacct->ac_stime += stime;
625 pacct->ac_minflt += current->min_flt;
626 pacct->ac_majflt += current->maj_flt;
627 spin_unlock_irq(¤t->sighand->siglock);
630 static void slow_acct_process(struct pid_namespace *ns)
632 for ( ; ns; ns = ns->parent) {
633 struct bsd_acct_struct *acct = acct_get(&ns->bacct);
635 do_acct_process(acct);
636 mutex_unlock(&acct->lock);
645 * handles process accounting for an exiting task
647 void acct_process(void)
649 struct pid_namespace *ns;
652 * This loop is safe lockless, since current is still
653 * alive and holds its namespace, which in turn holds
656 for (ns = task_active_pid_ns(current); ns != NULL; ns = ns->parent) {
661 slow_acct_process(ns);