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 pr_info("Process accounting paused\n");
114 u64 resume = sbuf.f_blocks * RESUME;
116 if (sbuf.f_bavail >= resume) {
118 pr_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 if (file->f_op->flush)
158 file->f_op->flush(file, NULL);
160 complete(&acct->done);
163 static void acct_kill(struct bsd_acct_struct *acct,
164 struct bsd_acct_struct *new)
167 struct pid_namespace *ns = acct->ns;
168 do_acct_process(acct);
169 INIT_WORK(&acct->work, close_work);
170 init_completion(&acct->done);
171 schedule_work(&acct->work);
172 wait_for_completion(&acct->done);
173 pin_remove(&acct->pin);
176 atomic_long_dec(&acct->pin.count);
177 mutex_unlock(&acct->lock);
182 static void acct_pin_kill(struct fs_pin *pin)
184 struct bsd_acct_struct *acct;
185 acct = container_of(pin, struct bsd_acct_struct, pin);
186 mutex_lock(&acct->lock);
188 mutex_unlock(&acct->lock);
192 acct_kill(acct, NULL);
195 static int acct_on(struct filename *pathname)
198 struct vfsmount *mnt, *internal;
199 struct pid_namespace *ns = task_active_pid_ns(current);
200 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);
225 internal = mnt_clone_internal(&file->f_path);
226 if (IS_ERR(internal)) {
228 filp_close(file, NULL);
229 return PTR_ERR(internal);
231 err = mnt_want_write(internal);
235 filp_close(file, NULL);
238 mnt = file->f_path.mnt;
239 file->f_path.mnt = internal;
241 atomic_long_set(&acct->pin.count, 1);
242 acct->pin.kill = acct_pin_kill;
244 acct->needcheck = jiffies;
246 mutex_init(&acct->lock);
247 mutex_lock_nested(&acct->lock, 1); /* nobody has seen it yet */
248 pin_insert(&acct->pin, mnt);
252 acct_kill(old, acct);
255 mutex_unlock(&acct->lock);
261 static DEFINE_MUTEX(acct_on_mutex);
264 * sys_acct - enable/disable process accounting
265 * @name: file name for accounting records or NULL to shutdown accounting
267 * Returns 0 for success or negative errno values for failure.
269 * sys_acct() is the only system call needed to implement process
270 * accounting. It takes the name of the file where accounting records
271 * should be written. If the filename is NULL, accounting will be
274 SYSCALL_DEFINE1(acct, const char __user *, name)
278 if (!capable(CAP_SYS_PACCT))
282 struct filename *tmp = getname(name);
286 mutex_lock(&acct_on_mutex);
287 error = acct_on(tmp);
288 mutex_unlock(&acct_on_mutex);
291 acct_kill(acct_get(task_active_pid_ns(current)), NULL);
297 void acct_exit_ns(struct pid_namespace *ns)
299 acct_kill(acct_get(ns), NULL);
303 * encode an unsigned long into a comp_t
305 * This routine has been adopted from the encode_comp_t() function in
306 * the kern_acct.c file of the FreeBSD operating system. The encoding
307 * is a 13-bit fraction with a 3-bit (base 8) exponent.
310 #define MANTSIZE 13 /* 13 bit mantissa. */
311 #define EXPSIZE 3 /* Base 8 (3 bit) exponent. */
312 #define MAXFRACT ((1 << MANTSIZE) - 1) /* Maximum fractional value. */
314 static comp_t encode_comp_t(unsigned long value)
319 while (value > MAXFRACT) {
320 rnd = value & (1 << (EXPSIZE - 1)); /* Round up? */
321 value >>= EXPSIZE; /* Base 8 exponent == 3 bit shift. */
326 * If we need to round up, do it (and handle overflow correctly).
328 if (rnd && (++value > MAXFRACT)) {
334 * Clean it up and polish it off.
336 exp <<= MANTSIZE; /* Shift the exponent into place */
337 exp += value; /* and add on the mantissa. */
341 #if ACCT_VERSION == 1 || ACCT_VERSION == 2
343 * encode an u64 into a comp2_t (24 bits)
345 * Format: 5 bit base 2 exponent, 20 bits mantissa.
346 * The leading bit of the mantissa is not stored, but implied for
347 * non-zero exponents.
348 * Largest encodable value is 50 bits.
351 #define MANTSIZE2 20 /* 20 bit mantissa. */
352 #define EXPSIZE2 5 /* 5 bit base 2 exponent. */
353 #define MAXFRACT2 ((1ul << MANTSIZE2) - 1) /* Maximum fractional value. */
354 #define MAXEXP2 ((1 << EXPSIZE2) - 1) /* Maximum exponent. */
356 static comp2_t encode_comp2_t(u64 value)
360 exp = (value > (MAXFRACT2>>1));
362 while (value > MAXFRACT2) {
369 * If we need to round up, do it (and handle overflow correctly).
371 if (rnd && (++value > MAXFRACT2)) {
377 /* Overflow. Return largest representable number instead. */
378 return (1ul << (MANTSIZE2+EXPSIZE2-1)) - 1;
380 return (value & (MAXFRACT2>>1)) | (exp << (MANTSIZE2-1));
385 #if ACCT_VERSION == 3
387 * encode an u64 into a 32 bit IEEE float
389 static u32 encode_float(u64 value)
396 while ((s64)value > 0) {
400 u = (u32)(value >> 40) & 0x7fffffu;
401 return u | (exp << 23);
406 * Write an accounting entry for an exiting process
408 * The acct_process() call is the workhorse of the process
409 * accounting system. The struct acct is built here and then written
410 * into the accounting file. This function should only be called from
411 * do_exit() or when switching to a different output file.
414 static void fill_ac(acct_t *ac)
416 struct pacct_struct *pacct = ¤t->signal->pacct;
417 u64 elapsed, run_time;
418 struct tty_struct *tty;
421 * Fill the accounting struct with the needed info as recorded
422 * by the different kernel functions.
424 memset(ac, 0, sizeof(acct_t));
426 ac->ac_version = ACCT_VERSION | ACCT_BYTEORDER;
427 strlcpy(ac->ac_comm, current->comm, sizeof(ac->ac_comm));
429 /* calculate run_time in nsec*/
430 run_time = ktime_get_ns();
431 run_time -= current->group_leader->start_time;
432 /* convert nsec -> AHZ */
433 elapsed = nsec_to_AHZ(run_time);
434 #if ACCT_VERSION == 3
435 ac->ac_etime = encode_float(elapsed);
437 ac->ac_etime = encode_comp_t(elapsed < (unsigned long) -1l ?
438 (unsigned long) elapsed : (unsigned long) -1l);
440 #if ACCT_VERSION == 1 || ACCT_VERSION == 2
442 /* new enlarged etime field */
443 comp2_t etime = encode_comp2_t(elapsed);
445 ac->ac_etime_hi = etime >> 16;
446 ac->ac_etime_lo = (u16) etime;
449 do_div(elapsed, AHZ);
450 ac->ac_btime = get_seconds() - elapsed;
455 spin_lock_irq(¤t->sighand->siglock);
456 tty = current->signal->tty; /* Safe as we hold the siglock */
457 ac->ac_tty = tty ? old_encode_dev(tty_devnum(tty)) : 0;
458 ac->ac_utime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_utime)));
459 ac->ac_stime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_stime)));
460 ac->ac_flag = pacct->ac_flag;
461 ac->ac_mem = encode_comp_t(pacct->ac_mem);
462 ac->ac_minflt = encode_comp_t(pacct->ac_minflt);
463 ac->ac_majflt = encode_comp_t(pacct->ac_majflt);
464 ac->ac_exitcode = pacct->ac_exitcode;
465 spin_unlock_irq(¤t->sighand->siglock);
468 * do_acct_process does all actual work. Caller holds the reference to file.
470 static void do_acct_process(struct bsd_acct_struct *acct)
474 const struct cred *orig_cred;
475 struct pid_namespace *ns = acct->ns;
476 struct file *file = acct->file;
479 * Accounting records are not subject to resource limits.
481 flim = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
482 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
483 /* Perform file operations on behalf of whoever enabled accounting */
484 orig_cred = override_creds(file->f_cred);
487 * First check to see if there is enough free_space to continue
488 * the process accounting system.
490 if (!check_free_space(acct))
494 /* we really need to bite the bullet and change layout */
495 ac.ac_uid = from_kuid_munged(file->f_cred->user_ns, orig_cred->uid);
496 ac.ac_gid = from_kgid_munged(file->f_cred->user_ns, orig_cred->gid);
497 #if ACCT_VERSION == 1 || ACCT_VERSION == 2
498 /* backward-compatible 16 bit fields */
499 ac.ac_uid16 = ac.ac_uid;
500 ac.ac_gid16 = ac.ac_gid;
502 #if ACCT_VERSION == 3
503 ac.ac_pid = task_tgid_nr_ns(current, ns);
505 ac.ac_ppid = task_tgid_nr_ns(rcu_dereference(current->real_parent), ns);
509 * Get freeze protection. If the fs is frozen, just skip the write
510 * as we could deadlock the system otherwise.
512 if (file_start_write_trylock(file)) {
513 /* it's been opened O_APPEND, so position is irrelevant */
515 __kernel_write(file, (char *)&ac, sizeof(acct_t), &pos);
516 file_end_write(file);
519 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = flim;
520 revert_creds(orig_cred);
524 * acct_collect - collect accounting information into pacct_struct
525 * @exitcode: task exit code
526 * @group_dead: not 0, if this thread is the last one in the process.
528 void acct_collect(long exitcode, int group_dead)
530 struct pacct_struct *pacct = ¤t->signal->pacct;
531 cputime_t utime, stime;
532 unsigned long vsize = 0;
534 if (group_dead && current->mm) {
535 struct vm_area_struct *vma;
537 down_read(¤t->mm->mmap_sem);
538 vma = current->mm->mmap;
540 vsize += vma->vm_end - vma->vm_start;
543 up_read(¤t->mm->mmap_sem);
546 spin_lock_irq(¤t->sighand->siglock);
548 pacct->ac_mem = vsize / 1024;
549 if (thread_group_leader(current)) {
550 pacct->ac_exitcode = exitcode;
551 if (current->flags & PF_FORKNOEXEC)
552 pacct->ac_flag |= AFORK;
554 if (current->flags & PF_SUPERPRIV)
555 pacct->ac_flag |= ASU;
556 if (current->flags & PF_DUMPCORE)
557 pacct->ac_flag |= ACORE;
558 if (current->flags & PF_SIGNALED)
559 pacct->ac_flag |= AXSIG;
560 task_cputime(current, &utime, &stime);
561 pacct->ac_utime += utime;
562 pacct->ac_stime += stime;
563 pacct->ac_minflt += current->min_flt;
564 pacct->ac_majflt += current->maj_flt;
565 spin_unlock_irq(¤t->sighand->siglock);
568 static void slow_acct_process(struct pid_namespace *ns)
570 for ( ; ns; ns = ns->parent) {
571 struct bsd_acct_struct *acct = acct_get(ns);
573 do_acct_process(acct);
574 mutex_unlock(&acct->lock);
583 * handles process accounting for an exiting task
585 void acct_process(void)
587 struct pid_namespace *ns;
590 * This loop is safe lockless, since current is still
591 * alive and holds its namespace, which in turn holds
594 for (ns = task_active_pid_ns(current); ns != NULL; ns = ns->parent) {
599 slow_acct_process(ns);