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 <linux/fs_pin.h>
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);
81 struct bsd_acct_struct {
85 unsigned long needcheck;
87 struct pid_namespace *ns;
88 struct work_struct work;
89 struct completion done;
93 * Check the amount of free space and suspend/resume accordingly.
95 static int check_free_space(struct bsd_acct_struct *acct)
99 if (time_is_before_jiffies(acct->needcheck))
103 if (vfs_statfs(&acct->file->f_path, &sbuf))
107 u64 suspend = sbuf.f_blocks * SUSPEND;
108 do_div(suspend, 100);
109 if (sbuf.f_bavail <= suspend) {
111 printk(KERN_INFO "Process accounting paused\n");
114 u64 resume = sbuf.f_blocks * RESUME;
116 if (sbuf.f_bavail >= resume) {
118 printk(KERN_INFO "Process accounting resumed\n");
122 acct->needcheck = jiffies + ACCT_TIMEOUT*HZ;
127 static struct bsd_acct_struct *acct_get(struct pid_namespace *ns)
129 struct bsd_acct_struct *res;
133 res = ACCESS_ONCE(ns->bacct);
138 if (!atomic_long_inc_not_zero(&res->pin.count)) {
144 mutex_lock(&res->lock);
146 mutex_unlock(&res->lock);
153 static void close_work(struct work_struct *work)
155 struct bsd_acct_struct *acct = container_of(work, struct bsd_acct_struct, work);
156 struct file *file = acct->file;
157 mnt_unpin(file->f_path.mnt);
158 if (file->f_op->flush)
159 file->f_op->flush(file, NULL);
161 complete(&acct->done);
164 static void acct_kill(struct bsd_acct_struct *acct,
165 struct bsd_acct_struct *new)
168 struct pid_namespace *ns = acct->ns;
169 do_acct_process(acct);
170 INIT_WORK(&acct->work, close_work);
171 init_completion(&acct->done);
172 schedule_work(&acct->work);
173 wait_for_completion(&acct->done);
174 pin_remove(&acct->pin);
177 atomic_long_dec(&acct->pin.count);
178 mutex_unlock(&acct->lock);
183 static void acct_pin_kill(struct fs_pin *pin)
185 struct bsd_acct_struct *acct;
186 acct = container_of(pin, struct bsd_acct_struct, pin);
187 mutex_lock(&acct->lock);
189 mutex_unlock(&acct->lock);
193 acct_kill(acct, NULL);
196 static int acct_on(struct filename *pathname)
199 struct vfsmount *mnt;
200 struct pid_namespace *ns = task_active_pid_ns(current);
201 struct bsd_acct_struct *acct, *old;
203 acct = kzalloc(sizeof(struct bsd_acct_struct), GFP_KERNEL);
207 /* Difference from BSD - they don't do O_APPEND */
208 file = file_open_name(pathname, O_WRONLY|O_APPEND|O_LARGEFILE, 0);
211 return PTR_ERR(file);
214 if (!S_ISREG(file_inode(file)->i_mode)) {
216 filp_close(file, NULL);
220 if (!file->f_op->write) {
222 filp_close(file, NULL);
226 atomic_long_set(&acct->pin.count, 1);
227 acct->pin.kill = acct_pin_kill;
229 acct->needcheck = jiffies;
231 mutex_init(&acct->lock);
232 mnt = file->f_path.mnt;
234 mutex_lock_nested(&acct->lock, 1); /* nobody has seen it yet */
235 pin_insert(&acct->pin, mnt);
239 acct_kill(old, acct);
242 mutex_unlock(&acct->lock);
243 mntput(mnt); /* it's pinned, now give up active reference */
247 static DEFINE_MUTEX(acct_on_mutex);
250 * sys_acct - enable/disable process accounting
251 * @name: file name for accounting records or NULL to shutdown accounting
253 * Returns 0 for success or negative errno values for failure.
255 * sys_acct() is the only system call needed to implement process
256 * accounting. It takes the name of the file where accounting records
257 * should be written. If the filename is NULL, accounting will be
260 SYSCALL_DEFINE1(acct, const char __user *, name)
264 if (!capable(CAP_SYS_PACCT))
268 struct filename *tmp = getname(name);
271 mutex_lock(&acct_on_mutex);
272 error = acct_on(tmp);
273 mutex_unlock(&acct_on_mutex);
276 acct_kill(acct_get(task_active_pid_ns(current)), NULL);
282 void acct_exit_ns(struct pid_namespace *ns)
284 acct_kill(acct_get(ns), NULL);
288 * encode an unsigned long into a comp_t
290 * This routine has been adopted from the encode_comp_t() function in
291 * the kern_acct.c file of the FreeBSD operating system. The encoding
292 * is a 13-bit fraction with a 3-bit (base 8) exponent.
295 #define MANTSIZE 13 /* 13 bit mantissa. */
296 #define EXPSIZE 3 /* Base 8 (3 bit) exponent. */
297 #define MAXFRACT ((1 << MANTSIZE) - 1) /* Maximum fractional value. */
299 static comp_t encode_comp_t(unsigned long value)
304 while (value > MAXFRACT) {
305 rnd = value & (1 << (EXPSIZE - 1)); /* Round up? */
306 value >>= EXPSIZE; /* Base 8 exponent == 3 bit shift. */
311 * If we need to round up, do it (and handle overflow correctly).
313 if (rnd && (++value > MAXFRACT)) {
319 * Clean it up and polish it off.
321 exp <<= MANTSIZE; /* Shift the exponent into place */
322 exp += value; /* and add on the mantissa. */
326 #if ACCT_VERSION==1 || ACCT_VERSION==2
328 * encode an u64 into a comp2_t (24 bits)
330 * Format: 5 bit base 2 exponent, 20 bits mantissa.
331 * The leading bit of the mantissa is not stored, but implied for
332 * non-zero exponents.
333 * Largest encodable value is 50 bits.
336 #define MANTSIZE2 20 /* 20 bit mantissa. */
337 #define EXPSIZE2 5 /* 5 bit base 2 exponent. */
338 #define MAXFRACT2 ((1ul << MANTSIZE2) - 1) /* Maximum fractional value. */
339 #define MAXEXP2 ((1 <<EXPSIZE2) - 1) /* Maximum exponent. */
341 static comp2_t encode_comp2_t(u64 value)
345 exp = (value > (MAXFRACT2>>1));
347 while (value > MAXFRACT2) {
354 * If we need to round up, do it (and handle overflow correctly).
356 if (rnd && (++value > MAXFRACT2)) {
362 /* Overflow. Return largest representable number instead. */
363 return (1ul << (MANTSIZE2+EXPSIZE2-1)) - 1;
365 return (value & (MAXFRACT2>>1)) | (exp << (MANTSIZE2-1));
372 * encode an u64 into a 32 bit IEEE float
374 static u32 encode_float(u64 value)
379 if (value==0) return 0;
380 while ((s64)value > 0){
384 u = (u32)(value >> 40) & 0x7fffffu;
385 return u | (exp << 23);
390 * Write an accounting entry for an exiting process
392 * The acct_process() call is the workhorse of the process
393 * accounting system. The struct acct is built here and then written
394 * into the accounting file. This function should only be called from
395 * do_exit() or when switching to a different output file.
398 static void fill_ac(acct_t *ac)
400 struct pacct_struct *pacct = ¤t->signal->pacct;
401 u64 elapsed, run_time;
402 struct tty_struct *tty;
405 * Fill the accounting struct with the needed info as recorded
406 * by the different kernel functions.
408 memset(ac, 0, sizeof(acct_t));
410 ac->ac_version = ACCT_VERSION | ACCT_BYTEORDER;
411 strlcpy(ac->ac_comm, current->comm, sizeof(ac->ac_comm));
413 /* calculate run_time in nsec*/
414 run_time = ktime_get_ns();
415 run_time -= current->group_leader->start_time;
416 /* convert nsec -> AHZ */
417 elapsed = nsec_to_AHZ(run_time);
419 ac->ac_etime = encode_float(elapsed);
421 ac->ac_etime = encode_comp_t(elapsed < (unsigned long) -1l ?
422 (unsigned long) elapsed : (unsigned long) -1l);
424 #if ACCT_VERSION==1 || ACCT_VERSION==2
426 /* new enlarged etime field */
427 comp2_t etime = encode_comp2_t(elapsed);
428 ac->ac_etime_hi = etime >> 16;
429 ac->ac_etime_lo = (u16) etime;
432 do_div(elapsed, AHZ);
433 ac->ac_btime = get_seconds() - elapsed;
438 spin_lock_irq(¤t->sighand->siglock);
439 tty = current->signal->tty; /* Safe as we hold the siglock */
440 ac->ac_tty = tty ? old_encode_dev(tty_devnum(tty)) : 0;
441 ac->ac_utime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_utime)));
442 ac->ac_stime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_stime)));
443 ac->ac_flag = pacct->ac_flag;
444 ac->ac_mem = encode_comp_t(pacct->ac_mem);
445 ac->ac_minflt = encode_comp_t(pacct->ac_minflt);
446 ac->ac_majflt = encode_comp_t(pacct->ac_majflt);
447 ac->ac_exitcode = pacct->ac_exitcode;
448 spin_unlock_irq(¤t->sighand->siglock);
451 * do_acct_process does all actual work. Caller holds the reference to file.
453 static void do_acct_process(struct bsd_acct_struct *acct)
457 const struct cred *orig_cred;
458 struct pid_namespace *ns = acct->ns;
459 struct file *file = acct->file;
462 * Accounting records are not subject to resource limits.
464 flim = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
465 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
466 /* Perform file operations on behalf of whoever enabled accounting */
467 orig_cred = override_creds(file->f_cred);
470 * First check to see if there is enough free_space to continue
471 * the process accounting system.
473 if (!check_free_space(acct))
477 /* we really need to bite the bullet and change layout */
478 ac.ac_uid = from_kuid_munged(file->f_cred->user_ns, orig_cred->uid);
479 ac.ac_gid = from_kgid_munged(file->f_cred->user_ns, orig_cred->gid);
480 #if ACCT_VERSION==1 || ACCT_VERSION==2
481 /* backward-compatible 16 bit fields */
482 ac.ac_uid16 = ac.ac_uid;
483 ac.ac_gid16 = ac.ac_gid;
486 ac.ac_pid = task_tgid_nr_ns(current, ns);
488 ac.ac_ppid = task_tgid_nr_ns(rcu_dereference(current->real_parent), ns);
492 * Get freeze protection. If the fs is frozen, just skip the write
493 * as we could deadlock the system otherwise.
495 if (file_start_write_trylock(file)) {
496 /* it's been opened O_APPEND, so position is irrelevant */
498 __kernel_write(file, (char *)&ac, sizeof(acct_t), &pos);
499 file_end_write(file);
502 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = flim;
503 revert_creds(orig_cred);
507 * acct_collect - collect accounting information into pacct_struct
508 * @exitcode: task exit code
509 * @group_dead: not 0, if this thread is the last one in the process.
511 void acct_collect(long exitcode, int group_dead)
513 struct pacct_struct *pacct = ¤t->signal->pacct;
514 cputime_t utime, stime;
515 unsigned long vsize = 0;
517 if (group_dead && current->mm) {
518 struct vm_area_struct *vma;
519 down_read(¤t->mm->mmap_sem);
520 vma = current->mm->mmap;
522 vsize += vma->vm_end - vma->vm_start;
525 up_read(¤t->mm->mmap_sem);
528 spin_lock_irq(¤t->sighand->siglock);
530 pacct->ac_mem = vsize / 1024;
531 if (thread_group_leader(current)) {
532 pacct->ac_exitcode = exitcode;
533 if (current->flags & PF_FORKNOEXEC)
534 pacct->ac_flag |= AFORK;
536 if (current->flags & PF_SUPERPRIV)
537 pacct->ac_flag |= ASU;
538 if (current->flags & PF_DUMPCORE)
539 pacct->ac_flag |= ACORE;
540 if (current->flags & PF_SIGNALED)
541 pacct->ac_flag |= AXSIG;
542 task_cputime(current, &utime, &stime);
543 pacct->ac_utime += utime;
544 pacct->ac_stime += stime;
545 pacct->ac_minflt += current->min_flt;
546 pacct->ac_majflt += current->maj_flt;
547 spin_unlock_irq(¤t->sighand->siglock);
550 static void slow_acct_process(struct pid_namespace *ns)
552 for ( ; ns; ns = ns->parent) {
553 struct bsd_acct_struct *acct = acct_get(ns);
555 do_acct_process(acct);
556 mutex_unlock(&acct->lock);
565 * handles process accounting for an exiting task
567 void acct_process(void)
569 struct pid_namespace *ns;
572 * This loop is safe lockless, since current is still
573 * alive and holds its namespace, which in turn holds
576 for (ns = task_active_pid_ns(current); ns != NULL; ns = ns->parent) {
581 slow_acct_process(ns);