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,
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 unsigned long needcheck;
90 struct pid_namespace *ns;
91 struct list_head list;
94 static DEFINE_SPINLOCK(acct_lock);
95 static LIST_HEAD(acct_list);
98 * Check the amount of free space and suspend/resume accordingly.
100 static int check_free_space(struct bsd_acct_struct *acct, struct file *file)
108 spin_lock(&acct_lock);
110 if (!file || time_is_before_jiffies(acct->needcheck))
112 spin_unlock(&acct_lock);
115 if (vfs_statfs(&file->f_path, &sbuf))
117 suspend = sbuf.f_blocks * SUSPEND;
118 resume = sbuf.f_blocks * RESUME;
120 do_div(suspend, 100);
123 if (sbuf.f_bavail <= suspend)
125 else if (sbuf.f_bavail >= resume)
131 * If some joker switched acct->file under us we'ld better be
132 * silent and _not_ touch anything.
134 spin_lock(&acct_lock);
135 if (file != acct->file) {
144 printk(KERN_INFO "Process accounting paused\n");
149 printk(KERN_INFO "Process accounting resumed\n");
153 acct->needcheck = jiffies + ACCT_TIMEOUT*HZ;
156 spin_unlock(&acct_lock);
161 * Close the old accounting file (if currently open) and then replace
162 * it with file (if non-NULL).
164 * NOTE: acct_lock MUST be held on entry and exit.
166 static void acct_file_reopen(struct bsd_acct_struct *acct, struct file *file,
167 struct pid_namespace *ns)
169 struct file *old_acct = NULL;
170 struct pid_namespace *old_ns = NULL;
173 old_acct = acct->file;
178 list_del(&acct->list);
183 acct->needcheck = jiffies + ACCT_TIMEOUT*HZ;
185 list_add(&acct->list, &acct_list);
188 mnt_unpin(old_acct->f_path.mnt);
189 spin_unlock(&acct_lock);
190 do_acct_process(acct, old_ns, old_acct);
191 filp_close(old_acct, NULL);
192 spin_lock(&acct_lock);
196 static int acct_on(struct filename *pathname)
199 struct vfsmount *mnt;
200 struct pid_namespace *ns;
201 struct bsd_acct_struct *acct = NULL;
203 /* Difference from BSD - they don't do O_APPEND */
204 file = file_open_name(pathname, O_WRONLY|O_APPEND|O_LARGEFILE, 0);
206 return PTR_ERR(file);
208 if (!S_ISREG(file_inode(file)->i_mode)) {
209 filp_close(file, NULL);
213 if (!file->f_op->write) {
214 filp_close(file, NULL);
218 ns = task_active_pid_ns(current);
219 if (ns->bacct == NULL) {
220 acct = kzalloc(sizeof(struct bsd_acct_struct), GFP_KERNEL);
222 filp_close(file, NULL);
227 spin_lock(&acct_lock);
228 if (ns->bacct == NULL) {
233 mnt = file->f_path.mnt;
235 acct_file_reopen(ns->bacct, file, ns);
236 spin_unlock(&acct_lock);
238 mntput(mnt); /* it's pinned, now give up active reference */
245 * sys_acct - enable/disable process accounting
246 * @name: file name for accounting records or NULL to shutdown accounting
248 * Returns 0 for success or negative errno values for failure.
250 * sys_acct() is the only system call needed to implement process
251 * accounting. It takes the name of the file where accounting records
252 * should be written. If the filename is NULL, accounting will be
255 SYSCALL_DEFINE1(acct, const char __user *, name)
259 if (!capable(CAP_SYS_PACCT))
263 struct filename *tmp = getname(name);
266 error = acct_on(tmp);
269 struct bsd_acct_struct *acct;
271 acct = task_active_pid_ns(current)->bacct;
275 spin_lock(&acct_lock);
276 acct_file_reopen(acct, NULL, NULL);
277 spin_unlock(&acct_lock);
284 * acct_auto_close - turn off a filesystem's accounting if it is on
285 * @m: vfsmount being shut down
287 * If the accounting is turned on for a file in the subtree pointed to
288 * to by m, turn accounting off. Done when m is about to die.
290 void acct_auto_close_mnt(struct vfsmount *m)
292 struct bsd_acct_struct *acct;
294 spin_lock(&acct_lock);
296 list_for_each_entry(acct, &acct_list, list)
297 if (acct->file && acct->file->f_path.mnt == m) {
298 acct_file_reopen(acct, NULL, NULL);
301 spin_unlock(&acct_lock);
305 * acct_auto_close - turn off a filesystem's accounting if it is on
306 * @sb: super block for the filesystem
308 * If the accounting is turned on for a file in the filesystem pointed
309 * to by sb, turn accounting off.
311 void acct_auto_close(struct super_block *sb)
313 struct bsd_acct_struct *acct;
315 spin_lock(&acct_lock);
317 list_for_each_entry(acct, &acct_list, list)
318 if (acct->file && acct->file->f_path.dentry->d_sb == sb) {
319 acct_file_reopen(acct, NULL, NULL);
322 spin_unlock(&acct_lock);
325 void acct_exit_ns(struct pid_namespace *ns)
327 struct bsd_acct_struct *acct = ns->bacct;
332 spin_lock(&acct_lock);
333 if (acct->file != NULL)
334 acct_file_reopen(acct, NULL, NULL);
335 spin_unlock(&acct_lock);
341 * encode an unsigned long into a comp_t
343 * This routine has been adopted from the encode_comp_t() function in
344 * the kern_acct.c file of the FreeBSD operating system. The encoding
345 * is a 13-bit fraction with a 3-bit (base 8) exponent.
348 #define MANTSIZE 13 /* 13 bit mantissa. */
349 #define EXPSIZE 3 /* Base 8 (3 bit) exponent. */
350 #define MAXFRACT ((1 << MANTSIZE) - 1) /* Maximum fractional value. */
352 static comp_t encode_comp_t(unsigned long value)
357 while (value > MAXFRACT) {
358 rnd = value & (1 << (EXPSIZE - 1)); /* Round up? */
359 value >>= EXPSIZE; /* Base 8 exponent == 3 bit shift. */
364 * If we need to round up, do it (and handle overflow correctly).
366 if (rnd && (++value > MAXFRACT)) {
372 * Clean it up and polish it off.
374 exp <<= MANTSIZE; /* Shift the exponent into place */
375 exp += value; /* and add on the mantissa. */
379 #if ACCT_VERSION==1 || ACCT_VERSION==2
381 * encode an u64 into a comp2_t (24 bits)
383 * Format: 5 bit base 2 exponent, 20 bits mantissa.
384 * The leading bit of the mantissa is not stored, but implied for
385 * non-zero exponents.
386 * Largest encodable value is 50 bits.
389 #define MANTSIZE2 20 /* 20 bit mantissa. */
390 #define EXPSIZE2 5 /* 5 bit base 2 exponent. */
391 #define MAXFRACT2 ((1ul << MANTSIZE2) - 1) /* Maximum fractional value. */
392 #define MAXEXP2 ((1 <<EXPSIZE2) - 1) /* Maximum exponent. */
394 static comp2_t encode_comp2_t(u64 value)
398 exp = (value > (MAXFRACT2>>1));
400 while (value > MAXFRACT2) {
407 * If we need to round up, do it (and handle overflow correctly).
409 if (rnd && (++value > MAXFRACT2)) {
415 /* Overflow. Return largest representable number instead. */
416 return (1ul << (MANTSIZE2+EXPSIZE2-1)) - 1;
418 return (value & (MAXFRACT2>>1)) | (exp << (MANTSIZE2-1));
425 * encode an u64 into a 32 bit IEEE float
427 static u32 encode_float(u64 value)
432 if (value==0) return 0;
433 while ((s64)value > 0){
437 u = (u32)(value >> 40) & 0x7fffffu;
438 return u | (exp << 23);
443 * Write an accounting entry for an exiting process
445 * The acct_process() call is the workhorse of the process
446 * accounting system. The struct acct is built here and then written
447 * into the accounting file. This function should only be called from
448 * do_exit() or when switching to a different output file.
452 * do_acct_process does all actual work. Caller holds the reference to file.
454 static void do_acct_process(struct bsd_acct_struct *acct,
455 struct pid_namespace *ns, struct file *file)
457 struct pacct_struct *pacct = ¤t->signal->pacct;
463 struct timespec uptime;
464 struct tty_struct *tty;
465 const struct cred *orig_cred;
467 /* Perform file operations on behalf of whoever enabled accounting */
468 orig_cred = override_creds(file->f_cred);
471 * First check to see if there is enough free_space to continue
472 * the process accounting system.
474 if (!check_free_space(acct, file))
478 * Fill the accounting struct with the needed info as recorded
479 * by the different kernel functions.
481 memset(&ac, 0, sizeof(acct_t));
483 ac.ac_version = ACCT_VERSION | ACCT_BYTEORDER;
484 strlcpy(ac.ac_comm, current->comm, sizeof(ac.ac_comm));
486 /* calculate run_time in nsec*/
487 do_posix_clock_monotonic_gettime(&uptime);
488 run_time = (u64)uptime.tv_sec*NSEC_PER_SEC + uptime.tv_nsec;
489 run_time -= (u64)current->group_leader->start_time.tv_sec * NSEC_PER_SEC
490 + current->group_leader->start_time.tv_nsec;
491 /* convert nsec -> AHZ */
492 elapsed = nsec_to_AHZ(run_time);
494 ac.ac_etime = encode_float(elapsed);
496 ac.ac_etime = encode_comp_t(elapsed < (unsigned long) -1l ?
497 (unsigned long) elapsed : (unsigned long) -1l);
499 #if ACCT_VERSION==1 || ACCT_VERSION==2
501 /* new enlarged etime field */
502 comp2_t etime = encode_comp2_t(elapsed);
503 ac.ac_etime_hi = etime >> 16;
504 ac.ac_etime_lo = (u16) etime;
507 do_div(elapsed, AHZ);
508 ac.ac_btime = get_seconds() - elapsed;
509 /* we really need to bite the bullet and change layout */
510 ac.ac_uid = from_kuid_munged(file->f_cred->user_ns, orig_cred->uid);
511 ac.ac_gid = from_kgid_munged(file->f_cred->user_ns, orig_cred->gid);
515 #if ACCT_VERSION==1 || ACCT_VERSION==2
516 /* backward-compatible 16 bit fields */
517 ac.ac_uid16 = ac.ac_uid;
518 ac.ac_gid16 = ac.ac_gid;
521 ac.ac_pid = task_tgid_nr_ns(current, ns);
523 ac.ac_ppid = task_tgid_nr_ns(rcu_dereference(current->real_parent), ns);
527 spin_lock_irq(¤t->sighand->siglock);
528 tty = current->signal->tty; /* Safe as we hold the siglock */
529 ac.ac_tty = tty ? old_encode_dev(tty_devnum(tty)) : 0;
530 ac.ac_utime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_utime)));
531 ac.ac_stime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_stime)));
532 ac.ac_flag = pacct->ac_flag;
533 ac.ac_mem = encode_comp_t(pacct->ac_mem);
534 ac.ac_minflt = encode_comp_t(pacct->ac_minflt);
535 ac.ac_majflt = encode_comp_t(pacct->ac_majflt);
536 ac.ac_exitcode = pacct->ac_exitcode;
537 spin_unlock_irq(¤t->sighand->siglock);
538 ac.ac_io = encode_comp_t(0 /* current->io_usage */); /* %% */
539 ac.ac_rw = encode_comp_t(ac.ac_io / 1024);
540 ac.ac_swaps = encode_comp_t(0);
543 * Get freeze protection. If the fs is frozen, just skip the write
544 * as we could deadlock the system otherwise.
546 if (!file_start_write_trylock(file))
549 * Kernel segment override to datasegment and write it
550 * to the accounting file.
555 * Accounting records are not subject to resource limits.
557 flim = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
558 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
559 file->f_op->write(file, (char *)&ac,
560 sizeof(acct_t), &file->f_pos);
561 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = flim;
563 file_end_write(file);
565 revert_creds(orig_cred);
569 * acct_collect - collect accounting information into pacct_struct
570 * @exitcode: task exit code
571 * @group_dead: not 0, if this thread is the last one in the process.
573 void acct_collect(long exitcode, int group_dead)
575 struct pacct_struct *pacct = ¤t->signal->pacct;
576 cputime_t utime, stime;
577 unsigned long vsize = 0;
579 if (group_dead && current->mm) {
580 struct vm_area_struct *vma;
581 down_read(¤t->mm->mmap_sem);
582 vma = current->mm->mmap;
584 vsize += vma->vm_end - vma->vm_start;
587 up_read(¤t->mm->mmap_sem);
590 spin_lock_irq(¤t->sighand->siglock);
592 pacct->ac_mem = vsize / 1024;
593 if (thread_group_leader(current)) {
594 pacct->ac_exitcode = exitcode;
595 if (current->flags & PF_FORKNOEXEC)
596 pacct->ac_flag |= AFORK;
598 if (current->flags & PF_SUPERPRIV)
599 pacct->ac_flag |= ASU;
600 if (current->flags & PF_DUMPCORE)
601 pacct->ac_flag |= ACORE;
602 if (current->flags & PF_SIGNALED)
603 pacct->ac_flag |= AXSIG;
604 task_cputime(current, &utime, &stime);
605 pacct->ac_utime += utime;
606 pacct->ac_stime += stime;
607 pacct->ac_minflt += current->min_flt;
608 pacct->ac_majflt += current->maj_flt;
609 spin_unlock_irq(¤t->sighand->siglock);
612 static void acct_process_in_ns(struct pid_namespace *ns)
614 struct file *file = NULL;
615 struct bsd_acct_struct *acct;
619 * accelerate the common fastpath:
621 if (!acct || !acct->file)
624 spin_lock(&acct_lock);
626 if (unlikely(!file)) {
627 spin_unlock(&acct_lock);
631 spin_unlock(&acct_lock);
633 do_acct_process(acct, ns, file);
638 * acct_process - now just a wrapper around acct_process_in_ns,
639 * which in turn is a wrapper around do_acct_process.
641 * handles process accounting for an exiting task
643 void acct_process(void)
645 struct pid_namespace *ns;
648 * This loop is safe lockless, since current is still
649 * alive and holds its namespace, which in turn holds
652 for (ns = task_active_pid_ns(current); ns != NULL; ns = ns->parent)
653 acct_process_in_ns(ns);