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>
62 #include <../fs/mount.h> /* will go away when we refactor */
65 * These constants control the amount of freespace that suspend and
66 * resume the process accounting system, and the time delay between
68 * Turned into sysctl-controllable parameters. AV, 12/11/98
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 */
77 * External references and all of the globals.
79 static void do_acct_process(struct bsd_acct_struct *acct);
85 struct hlist_node s_list;
86 struct hlist_node m_list;
92 struct bsd_acct_struct {
96 unsigned long needcheck;
98 struct pid_namespace *ns;
99 struct work_struct work;
100 struct completion done;
103 static void pin_free_rcu(struct rcu_head *head)
105 kfree(container_of(head, struct fs_pin, rcu));
108 static DEFINE_SPINLOCK(acct_lock);
111 * Check the amount of free space and suspend/resume accordingly.
113 static int check_free_space(struct bsd_acct_struct *acct)
117 if (time_is_before_jiffies(acct->needcheck))
121 if (vfs_statfs(&acct->file->f_path, &sbuf))
125 u64 suspend = sbuf.f_blocks * SUSPEND;
126 do_div(suspend, 100);
127 if (sbuf.f_bavail <= suspend) {
129 printk(KERN_INFO "Process accounting paused\n");
132 u64 resume = sbuf.f_blocks * RESUME;
134 if (sbuf.f_bavail >= resume) {
136 printk(KERN_INFO "Process accounting resumed\n");
140 acct->needcheck = jiffies + ACCT_TIMEOUT*HZ;
145 static void pin_put(struct fs_pin *p)
147 if (atomic_long_dec_and_test(&p->count))
148 call_rcu(&p->rcu, pin_free_rcu);
151 static struct bsd_acct_struct *__acct_get(struct bsd_acct_struct *res)
153 if (!atomic_long_inc_not_zero(&res->pin.count)) {
159 mutex_lock(&res->lock);
161 mutex_unlock(&res->lock);
168 static struct bsd_acct_struct *acct_get(struct pid_namespace *ns)
170 struct bsd_acct_struct *res;
174 res = ACCESS_ONCE(ns->bacct);
179 res = __acct_get(res);
185 static void close_work(struct work_struct *work)
187 struct bsd_acct_struct *acct = container_of(work, struct bsd_acct_struct, work);
188 struct file *file = acct->file;
189 mnt_unpin(file->f_path.mnt);
190 if (file->f_op->flush)
191 file->f_op->flush(file, NULL);
193 complete(&acct->done);
196 static void acct_kill(struct bsd_acct_struct *acct,
197 struct bsd_acct_struct *new)
200 struct pid_namespace *ns = acct->ns;
201 do_acct_process(acct);
202 INIT_WORK(&acct->work, close_work);
203 init_completion(&acct->done);
204 schedule_work(&acct->work);
205 wait_for_completion(&acct->done);
206 spin_lock(&acct_lock);
207 hlist_del(&acct->pin.m_list);
208 hlist_del(&acct->pin.s_list);
209 spin_unlock(&acct_lock);
212 struct vfsmount *m = new->file->f_path.mnt;
213 spin_lock(&acct_lock);
214 hlist_add_head(&new->pin.s_list, &m->mnt_sb->s_pins);
215 hlist_add_head(&new->pin.m_list, &real_mount(m)->mnt_pins);
216 spin_unlock(&acct_lock);
217 mutex_unlock(&new->lock);
220 atomic_long_dec(&acct->pin.count);
221 mutex_unlock(&acct->lock);
226 static int acct_on(struct filename *pathname)
229 struct vfsmount *mnt;
230 struct pid_namespace *ns = task_active_pid_ns(current);
231 struct bsd_acct_struct *acct, *old;
233 acct = kzalloc(sizeof(struct bsd_acct_struct), GFP_KERNEL);
237 /* Difference from BSD - they don't do O_APPEND */
238 file = file_open_name(pathname, O_WRONLY|O_APPEND|O_LARGEFILE, 0);
241 return PTR_ERR(file);
244 if (!S_ISREG(file_inode(file)->i_mode)) {
246 filp_close(file, NULL);
250 if (!file->f_op->write) {
252 filp_close(file, NULL);
256 atomic_long_set(&acct->pin.count, 1);
258 acct->needcheck = jiffies;
260 mutex_init(&acct->lock);
261 mnt = file->f_path.mnt;
265 mutex_lock_nested(&acct->lock, 1); /* nobody has seen it yet */
267 acct_kill(old, acct);
270 spin_lock(&acct_lock);
271 hlist_add_head(&acct->pin.s_list, &mnt->mnt_sb->s_pins);
272 hlist_add_head(&acct->pin.m_list, &real_mount(mnt)->mnt_pins);
273 spin_unlock(&acct_lock);
274 mutex_unlock(&acct->lock);
276 mntput(mnt); /* it's pinned, now give up active reference */
280 static DEFINE_MUTEX(acct_on_mutex);
283 * sys_acct - enable/disable process accounting
284 * @name: file name for accounting records or NULL to shutdown accounting
286 * Returns 0 for success or negative errno values for failure.
288 * sys_acct() is the only system call needed to implement process
289 * accounting. It takes the name of the file where accounting records
290 * should be written. If the filename is NULL, accounting will be
293 SYSCALL_DEFINE1(acct, const char __user *, name)
297 if (!capable(CAP_SYS_PACCT))
301 struct filename *tmp = getname(name);
304 mutex_lock(&acct_on_mutex);
305 error = acct_on(tmp);
306 mutex_unlock(&acct_on_mutex);
309 acct_kill(acct_get(task_active_pid_ns(current)), NULL);
315 void acct_auto_close_mnt(struct hlist_head *list)
319 struct hlist_node *p = ACCESS_ONCE(list->first);
322 acct_kill(__acct_get(hlist_entry(p,
323 struct bsd_acct_struct,
330 void acct_auto_close(struct hlist_head *list)
334 struct hlist_node *p = ACCESS_ONCE(list->first);
337 acct_kill(__acct_get(hlist_entry(p,
338 struct bsd_acct_struct,
345 void acct_exit_ns(struct pid_namespace *ns)
347 acct_kill(acct_get(ns), NULL);
351 * encode an unsigned long into a comp_t
353 * This routine has been adopted from the encode_comp_t() function in
354 * the kern_acct.c file of the FreeBSD operating system. The encoding
355 * is a 13-bit fraction with a 3-bit (base 8) exponent.
358 #define MANTSIZE 13 /* 13 bit mantissa. */
359 #define EXPSIZE 3 /* Base 8 (3 bit) exponent. */
360 #define MAXFRACT ((1 << MANTSIZE) - 1) /* Maximum fractional value. */
362 static comp_t encode_comp_t(unsigned long value)
367 while (value > MAXFRACT) {
368 rnd = value & (1 << (EXPSIZE - 1)); /* Round up? */
369 value >>= EXPSIZE; /* Base 8 exponent == 3 bit shift. */
374 * If we need to round up, do it (and handle overflow correctly).
376 if (rnd && (++value > MAXFRACT)) {
382 * Clean it up and polish it off.
384 exp <<= MANTSIZE; /* Shift the exponent into place */
385 exp += value; /* and add on the mantissa. */
389 #if ACCT_VERSION==1 || ACCT_VERSION==2
391 * encode an u64 into a comp2_t (24 bits)
393 * Format: 5 bit base 2 exponent, 20 bits mantissa.
394 * The leading bit of the mantissa is not stored, but implied for
395 * non-zero exponents.
396 * Largest encodable value is 50 bits.
399 #define MANTSIZE2 20 /* 20 bit mantissa. */
400 #define EXPSIZE2 5 /* 5 bit base 2 exponent. */
401 #define MAXFRACT2 ((1ul << MANTSIZE2) - 1) /* Maximum fractional value. */
402 #define MAXEXP2 ((1 <<EXPSIZE2) - 1) /* Maximum exponent. */
404 static comp2_t encode_comp2_t(u64 value)
408 exp = (value > (MAXFRACT2>>1));
410 while (value > MAXFRACT2) {
417 * If we need to round up, do it (and handle overflow correctly).
419 if (rnd && (++value > MAXFRACT2)) {
425 /* Overflow. Return largest representable number instead. */
426 return (1ul << (MANTSIZE2+EXPSIZE2-1)) - 1;
428 return (value & (MAXFRACT2>>1)) | (exp << (MANTSIZE2-1));
435 * encode an u64 into a 32 bit IEEE float
437 static u32 encode_float(u64 value)
442 if (value==0) return 0;
443 while ((s64)value > 0){
447 u = (u32)(value >> 40) & 0x7fffffu;
448 return u | (exp << 23);
453 * Write an accounting entry for an exiting process
455 * The acct_process() call is the workhorse of the process
456 * accounting system. The struct acct is built here and then written
457 * into the accounting file. This function should only be called from
458 * do_exit() or when switching to a different output file.
461 static void fill_ac(acct_t *ac)
463 struct pacct_struct *pacct = ¤t->signal->pacct;
464 u64 elapsed, run_time;
465 struct tty_struct *tty;
468 * Fill the accounting struct with the needed info as recorded
469 * by the different kernel functions.
471 memset(ac, 0, sizeof(acct_t));
473 ac->ac_version = ACCT_VERSION | ACCT_BYTEORDER;
474 strlcpy(ac->ac_comm, current->comm, sizeof(ac->ac_comm));
476 /* calculate run_time in nsec*/
477 run_time = ktime_get_ns();
478 run_time -= current->group_leader->start_time;
479 /* convert nsec -> AHZ */
480 elapsed = nsec_to_AHZ(run_time);
482 ac->ac_etime = encode_float(elapsed);
484 ac->ac_etime = encode_comp_t(elapsed < (unsigned long) -1l ?
485 (unsigned long) elapsed : (unsigned long) -1l);
487 #if ACCT_VERSION==1 || ACCT_VERSION==2
489 /* new enlarged etime field */
490 comp2_t etime = encode_comp2_t(elapsed);
491 ac->ac_etime_hi = etime >> 16;
492 ac->ac_etime_lo = (u16) etime;
495 do_div(elapsed, AHZ);
496 ac->ac_btime = get_seconds() - elapsed;
501 spin_lock_irq(¤t->sighand->siglock);
502 tty = current->signal->tty; /* Safe as we hold the siglock */
503 ac->ac_tty = tty ? old_encode_dev(tty_devnum(tty)) : 0;
504 ac->ac_utime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_utime)));
505 ac->ac_stime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_stime)));
506 ac->ac_flag = pacct->ac_flag;
507 ac->ac_mem = encode_comp_t(pacct->ac_mem);
508 ac->ac_minflt = encode_comp_t(pacct->ac_minflt);
509 ac->ac_majflt = encode_comp_t(pacct->ac_majflt);
510 ac->ac_exitcode = pacct->ac_exitcode;
511 spin_unlock_irq(¤t->sighand->siglock);
514 * do_acct_process does all actual work. Caller holds the reference to file.
516 static void do_acct_process(struct bsd_acct_struct *acct)
520 const struct cred *orig_cred;
521 struct pid_namespace *ns = acct->ns;
522 struct file *file = acct->file;
525 * Accounting records are not subject to resource limits.
527 flim = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
528 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
529 /* Perform file operations on behalf of whoever enabled accounting */
530 orig_cred = override_creds(file->f_cred);
533 * First check to see if there is enough free_space to continue
534 * the process accounting system.
536 if (!check_free_space(acct))
540 /* we really need to bite the bullet and change layout */
541 ac.ac_uid = from_kuid_munged(file->f_cred->user_ns, orig_cred->uid);
542 ac.ac_gid = from_kgid_munged(file->f_cred->user_ns, orig_cred->gid);
543 #if ACCT_VERSION==1 || ACCT_VERSION==2
544 /* backward-compatible 16 bit fields */
545 ac.ac_uid16 = ac.ac_uid;
546 ac.ac_gid16 = ac.ac_gid;
549 ac.ac_pid = task_tgid_nr_ns(current, ns);
551 ac.ac_ppid = task_tgid_nr_ns(rcu_dereference(current->real_parent), ns);
555 * Get freeze protection. If the fs is frozen, just skip the write
556 * as we could deadlock the system otherwise.
558 if (file_start_write_trylock(file)) {
559 /* it's been opened O_APPEND, so position is irrelevant */
561 __kernel_write(file, (char *)&ac, sizeof(acct_t), &pos);
562 file_end_write(file);
565 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = flim;
566 revert_creds(orig_cred);
570 * acct_collect - collect accounting information into pacct_struct
571 * @exitcode: task exit code
572 * @group_dead: not 0, if this thread is the last one in the process.
574 void acct_collect(long exitcode, int group_dead)
576 struct pacct_struct *pacct = ¤t->signal->pacct;
577 cputime_t utime, stime;
578 unsigned long vsize = 0;
580 if (group_dead && current->mm) {
581 struct vm_area_struct *vma;
582 down_read(¤t->mm->mmap_sem);
583 vma = current->mm->mmap;
585 vsize += vma->vm_end - vma->vm_start;
588 up_read(¤t->mm->mmap_sem);
591 spin_lock_irq(¤t->sighand->siglock);
593 pacct->ac_mem = vsize / 1024;
594 if (thread_group_leader(current)) {
595 pacct->ac_exitcode = exitcode;
596 if (current->flags & PF_FORKNOEXEC)
597 pacct->ac_flag |= AFORK;
599 if (current->flags & PF_SUPERPRIV)
600 pacct->ac_flag |= ASU;
601 if (current->flags & PF_DUMPCORE)
602 pacct->ac_flag |= ACORE;
603 if (current->flags & PF_SIGNALED)
604 pacct->ac_flag |= AXSIG;
605 task_cputime(current, &utime, &stime);
606 pacct->ac_utime += utime;
607 pacct->ac_stime += stime;
608 pacct->ac_minflt += current->min_flt;
609 pacct->ac_majflt += current->maj_flt;
610 spin_unlock_irq(¤t->sighand->siglock);
613 static void slow_acct_process(struct pid_namespace *ns)
615 for ( ; ns; ns = ns->parent) {
616 struct bsd_acct_struct *acct = acct_get(ns);
618 do_acct_process(acct);
619 mutex_unlock(&acct->lock);
628 * handles process accounting for an exiting task
630 void acct_process(void)
632 struct pid_namespace *ns;
635 * This loop is safe lockless, since current is still
636 * alive and holds its namespace, which in turn holds
639 for (ns = task_active_pid_ns(current); ns != NULL; ns = ns->parent) {
644 slow_acct_process(ns);