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 <asm/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,
79 struct pid_namespace *ns, struct file *);
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.
86 struct bsd_acct_struct {
88 volatile int needcheck;
90 struct pid_namespace *ns;
91 struct timer_list timer;
92 struct list_head list;
95 static DEFINE_SPINLOCK(acct_lock);
96 static LIST_HEAD(acct_list);
99 * Called whenever the timer says to check the free space.
101 static void acct_timeout(unsigned long x)
103 struct bsd_acct_struct *acct = (struct bsd_acct_struct *)x;
108 * Check the amount of free space and suspend/resume accordingly.
110 static int check_free_space(struct bsd_acct_struct *acct, struct file *file)
118 spin_lock(&acct_lock);
120 if (!file || !acct->needcheck)
122 spin_unlock(&acct_lock);
125 if (vfs_statfs(file->f_path.dentry, &sbuf))
127 suspend = sbuf.f_blocks * SUSPEND;
128 resume = sbuf.f_blocks * RESUME;
130 sector_div(suspend, 100);
131 sector_div(resume, 100);
133 if (sbuf.f_bavail <= suspend)
135 else if (sbuf.f_bavail >= resume)
141 * If some joker switched acct->file under us we'ld better be
142 * silent and _not_ touch anything.
144 spin_lock(&acct_lock);
145 if (file != acct->file) {
154 printk(KERN_INFO "Process accounting paused\n");
159 printk(KERN_INFO "Process accounting resumed\n");
163 del_timer(&acct->timer);
165 acct->timer.expires = jiffies + ACCT_TIMEOUT*HZ;
166 add_timer(&acct->timer);
169 spin_unlock(&acct_lock);
174 * Close the old accounting file (if currently open) and then replace
175 * it with file (if non-NULL).
177 * NOTE: acct_lock MUST be held on entry and exit.
179 static void acct_file_reopen(struct bsd_acct_struct *acct, struct file *file,
180 struct pid_namespace *ns)
182 struct file *old_acct = NULL;
183 struct pid_namespace *old_ns = NULL;
186 old_acct = acct->file;
188 del_timer(&acct->timer);
193 list_del(&acct->list);
200 list_add(&acct->list, &acct_list);
201 /* It's been deleted if it was used before so this is safe */
202 setup_timer(&acct->timer, acct_timeout, (unsigned long)acct);
203 acct->timer.expires = jiffies + ACCT_TIMEOUT*HZ;
204 add_timer(&acct->timer);
207 mnt_unpin(old_acct->f_path.mnt);
208 spin_unlock(&acct_lock);
209 do_acct_process(acct, old_ns, old_acct);
210 filp_close(old_acct, NULL);
211 spin_lock(&acct_lock);
215 static int acct_on(char *name)
218 struct vfsmount *mnt;
219 struct pid_namespace *ns;
220 struct bsd_acct_struct *acct = NULL;
222 /* Difference from BSD - they don't do O_APPEND */
223 file = filp_open(name, O_WRONLY|O_APPEND|O_LARGEFILE, 0);
225 return PTR_ERR(file);
227 if (!S_ISREG(file->f_path.dentry->d_inode->i_mode)) {
228 filp_close(file, NULL);
232 if (!file->f_op->write) {
233 filp_close(file, NULL);
237 ns = task_active_pid_ns(current);
238 if (ns->bacct == NULL) {
239 acct = kzalloc(sizeof(struct bsd_acct_struct), GFP_KERNEL);
241 filp_close(file, NULL);
246 spin_lock(&acct_lock);
247 if (ns->bacct == NULL) {
252 mnt = file->f_path.mnt;
254 acct_file_reopen(ns->bacct, file, ns);
255 spin_unlock(&acct_lock);
257 mntput(mnt); /* it's pinned, now give up active reference */
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 char *tmp = getname(name);
284 return (PTR_ERR(tmp));
285 error = acct_on(tmp);
288 struct bsd_acct_struct *acct;
290 acct = task_active_pid_ns(current)->bacct;
294 spin_lock(&acct_lock);
295 acct_file_reopen(acct, NULL, NULL);
296 spin_unlock(&acct_lock);
303 * acct_auto_close - turn off a filesystem's accounting if it is on
304 * @m: vfsmount being shut down
306 * If the accounting is turned on for a file in the subtree pointed to
307 * to by m, turn accounting off. Done when m is about to die.
309 void acct_auto_close_mnt(struct vfsmount *m)
311 struct bsd_acct_struct *acct;
313 spin_lock(&acct_lock);
315 list_for_each_entry(acct, &acct_list, list)
316 if (acct->file && acct->file->f_path.mnt == m) {
317 acct_file_reopen(acct, NULL, NULL);
320 spin_unlock(&acct_lock);
324 * acct_auto_close - turn off a filesystem's accounting if it is on
325 * @sb: super block for the filesystem
327 * If the accounting is turned on for a file in the filesystem pointed
328 * to by sb, turn accounting off.
330 void acct_auto_close(struct super_block *sb)
332 struct bsd_acct_struct *acct;
334 spin_lock(&acct_lock);
336 list_for_each_entry(acct, &acct_list, list)
337 if (acct->file && acct->file->f_path.mnt->mnt_sb == sb) {
338 acct_file_reopen(acct, NULL, NULL);
341 spin_unlock(&acct_lock);
344 void acct_exit_ns(struct pid_namespace *ns)
346 struct bsd_acct_struct *acct = ns->bacct;
351 del_timer_sync(&acct->timer);
352 spin_lock(&acct_lock);
353 if (acct->file != NULL)
354 acct_file_reopen(acct, NULL, NULL);
355 spin_unlock(&acct_lock);
361 * encode an unsigned long into a comp_t
363 * This routine has been adopted from the encode_comp_t() function in
364 * the kern_acct.c file of the FreeBSD operating system. The encoding
365 * is a 13-bit fraction with a 3-bit (base 8) exponent.
368 #define MANTSIZE 13 /* 13 bit mantissa. */
369 #define EXPSIZE 3 /* Base 8 (3 bit) exponent. */
370 #define MAXFRACT ((1 << MANTSIZE) - 1) /* Maximum fractional value. */
372 static comp_t encode_comp_t(unsigned long value)
377 while (value > MAXFRACT) {
378 rnd = value & (1 << (EXPSIZE - 1)); /* Round up? */
379 value >>= EXPSIZE; /* Base 8 exponent == 3 bit shift. */
384 * If we need to round up, do it (and handle overflow correctly).
386 if (rnd && (++value > MAXFRACT)) {
392 * Clean it up and polish it off.
394 exp <<= MANTSIZE; /* Shift the exponent into place */
395 exp += value; /* and add on the mantissa. */
399 #if ACCT_VERSION==1 || ACCT_VERSION==2
401 * encode an u64 into a comp2_t (24 bits)
403 * Format: 5 bit base 2 exponent, 20 bits mantissa.
404 * The leading bit of the mantissa is not stored, but implied for
405 * non-zero exponents.
406 * Largest encodable value is 50 bits.
409 #define MANTSIZE2 20 /* 20 bit mantissa. */
410 #define EXPSIZE2 5 /* 5 bit base 2 exponent. */
411 #define MAXFRACT2 ((1ul << MANTSIZE2) - 1) /* Maximum fractional value. */
412 #define MAXEXP2 ((1 <<EXPSIZE2) - 1) /* Maximum exponent. */
414 static comp2_t encode_comp2_t(u64 value)
418 exp = (value > (MAXFRACT2>>1));
420 while (value > MAXFRACT2) {
427 * If we need to round up, do it (and handle overflow correctly).
429 if (rnd && (++value > MAXFRACT2)) {
435 /* Overflow. Return largest representable number instead. */
436 return (1ul << (MANTSIZE2+EXPSIZE2-1)) - 1;
438 return (value & (MAXFRACT2>>1)) | (exp << (MANTSIZE2-1));
445 * encode an u64 into a 32 bit IEEE float
447 static u32 encode_float(u64 value)
452 if (value==0) return 0;
453 while ((s64)value > 0){
457 u = (u32)(value >> 40) & 0x7fffffu;
458 return u | (exp << 23);
463 * Write an accounting entry for an exiting process
465 * The acct_process() call is the workhorse of the process
466 * accounting system. The struct acct is built here and then written
467 * into the accounting file. This function should only be called from
468 * do_exit() or when switching to a different output file.
472 * do_acct_process does all actual work. Caller holds the reference to file.
474 static void do_acct_process(struct bsd_acct_struct *acct,
475 struct pid_namespace *ns, struct file *file)
477 struct pacct_struct *pacct = ¤t->signal->pacct;
483 struct timespec uptime;
484 struct tty_struct *tty;
485 const struct cred *orig_cred;
487 /* Perform file operations on behalf of whoever enabled accounting */
488 orig_cred = override_creds(file->f_cred);
491 * First check to see if there is enough free_space to continue
492 * the process accounting system.
494 if (!check_free_space(acct, file))
498 * Fill the accounting struct with the needed info as recorded
499 * by the different kernel functions.
501 memset((caddr_t)&ac, 0, sizeof(acct_t));
503 ac.ac_version = ACCT_VERSION | ACCT_BYTEORDER;
504 strlcpy(ac.ac_comm, current->comm, sizeof(ac.ac_comm));
506 /* calculate run_time in nsec*/
507 do_posix_clock_monotonic_gettime(&uptime);
508 run_time = (u64)uptime.tv_sec*NSEC_PER_SEC + uptime.tv_nsec;
509 run_time -= (u64)current->group_leader->start_time.tv_sec * NSEC_PER_SEC
510 + current->group_leader->start_time.tv_nsec;
511 /* convert nsec -> AHZ */
512 elapsed = nsec_to_AHZ(run_time);
514 ac.ac_etime = encode_float(elapsed);
516 ac.ac_etime = encode_comp_t(elapsed < (unsigned long) -1l ?
517 (unsigned long) elapsed : (unsigned long) -1l);
519 #if ACCT_VERSION==1 || ACCT_VERSION==2
521 /* new enlarged etime field */
522 comp2_t etime = encode_comp2_t(elapsed);
523 ac.ac_etime_hi = etime >> 16;
524 ac.ac_etime_lo = (u16) etime;
527 do_div(elapsed, AHZ);
528 ac.ac_btime = get_seconds() - elapsed;
529 /* we really need to bite the bullet and change layout */
530 ac.ac_uid = orig_cred->uid;
531 ac.ac_gid = orig_cred->gid;
535 #if ACCT_VERSION==1 || ACCT_VERSION==2
536 /* backward-compatible 16 bit fields */
537 ac.ac_uid16 = ac.ac_uid;
538 ac.ac_gid16 = ac.ac_gid;
541 ac.ac_pid = task_tgid_nr_ns(current, ns);
543 ac.ac_ppid = task_tgid_nr_ns(rcu_dereference(current->real_parent), ns);
547 spin_lock_irq(¤t->sighand->siglock);
548 tty = current->signal->tty; /* Safe as we hold the siglock */
549 ac.ac_tty = tty ? old_encode_dev(tty_devnum(tty)) : 0;
550 ac.ac_utime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_utime)));
551 ac.ac_stime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_stime)));
552 ac.ac_flag = pacct->ac_flag;
553 ac.ac_mem = encode_comp_t(pacct->ac_mem);
554 ac.ac_minflt = encode_comp_t(pacct->ac_minflt);
555 ac.ac_majflt = encode_comp_t(pacct->ac_majflt);
556 ac.ac_exitcode = pacct->ac_exitcode;
557 spin_unlock_irq(¤t->sighand->siglock);
558 ac.ac_io = encode_comp_t(0 /* current->io_usage */); /* %% */
559 ac.ac_rw = encode_comp_t(ac.ac_io / 1024);
560 ac.ac_swaps = encode_comp_t(0);
563 * Kernel segment override to datasegment and write it
564 * to the accounting file.
569 * Accounting records are not subject to resource limits.
571 flim = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
572 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
573 file->f_op->write(file, (char *)&ac,
574 sizeof(acct_t), &file->f_pos);
575 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = flim;
578 revert_creds(orig_cred);
582 * acct_collect - collect accounting information into pacct_struct
583 * @exitcode: task exit code
584 * @group_dead: not 0, if this thread is the last one in the process.
586 void acct_collect(long exitcode, int group_dead)
588 struct pacct_struct *pacct = ¤t->signal->pacct;
589 unsigned long vsize = 0;
591 if (group_dead && current->mm) {
592 struct vm_area_struct *vma;
593 down_read(¤t->mm->mmap_sem);
594 vma = current->mm->mmap;
596 vsize += vma->vm_end - vma->vm_start;
599 up_read(¤t->mm->mmap_sem);
602 spin_lock_irq(¤t->sighand->siglock);
604 pacct->ac_mem = vsize / 1024;
605 if (thread_group_leader(current)) {
606 pacct->ac_exitcode = exitcode;
607 if (current->flags & PF_FORKNOEXEC)
608 pacct->ac_flag |= AFORK;
610 if (current->flags & PF_SUPERPRIV)
611 pacct->ac_flag |= ASU;
612 if (current->flags & PF_DUMPCORE)
613 pacct->ac_flag |= ACORE;
614 if (current->flags & PF_SIGNALED)
615 pacct->ac_flag |= AXSIG;
616 pacct->ac_utime = cputime_add(pacct->ac_utime, current->utime);
617 pacct->ac_stime = cputime_add(pacct->ac_stime, current->stime);
618 pacct->ac_minflt += current->min_flt;
619 pacct->ac_majflt += current->maj_flt;
620 spin_unlock_irq(¤t->sighand->siglock);
623 static void acct_process_in_ns(struct pid_namespace *ns)
625 struct file *file = NULL;
626 struct bsd_acct_struct *acct;
630 * accelerate the common fastpath:
632 if (!acct || !acct->file)
635 spin_lock(&acct_lock);
637 if (unlikely(!file)) {
638 spin_unlock(&acct_lock);
642 spin_unlock(&acct_lock);
644 do_acct_process(acct, ns, file);
649 * acct_process - now just a wrapper around acct_process_in_ns,
650 * which in turn is a wrapper around do_acct_process.
652 * handles process accounting for an exiting task
654 void acct_process(void)
656 struct pid_namespace *ns;
659 * This loop is safe lockless, since current is still
660 * alive and holds its namespace, which in turn holds
663 for (ns = task_active_pid_ns(current); ns != NULL; ns = ns->parent)
664 acct_process_in_ns(ns);