1 // SPDX-License-Identifier: GPL-2.0
5 * BSD Process Accounting for Linux
7 * Author: Marco van Wieringen <mvw@planets.elm.net>
9 * Some code based on ideas and code from:
10 * Thomas K. Dyas <tdyas@eden.rutgers.edu>
12 * This file implements BSD-style process accounting. Whenever any
13 * process exits, an accounting record of type "struct acct" is
14 * written to the file specified with the acct() system call. It is
15 * up to user-level programs to do useful things with the accounting
16 * log. The kernel just provides the raw accounting information.
18 * (C) Copyright 1995 - 1997 Marco van Wieringen - ELM Consultancy B.V.
20 * Plugged two leaks. 1) It didn't return acct_file into the free_filps if
21 * the file happened to be read-only. 2) If the accounting was suspended
22 * due to the lack of space it happily allowed to reopen it and completely
23 * lost the old acct_file. 3/10/98, Al Viro.
25 * Now we silently close acct_file on attempt to reopen. Cleaned sys_acct().
26 * XTerms and EMACS are manifestations of pure evil. 21/10/98, AV.
28 * Fixed a nasty interaction with sys_umount(). If the accounting
29 * was suspeneded we failed to stop it on umount(). Messy.
30 * Another one: remount to readonly didn't stop accounting.
31 * Question: what should we do if we have CAP_SYS_ADMIN but not
32 * CAP_SYS_PACCT? Current code does the following: umount returns -EBUSY
33 * unless we are messing with the root. In that case we are getting a
34 * real mess with do_remount_sb(). 9/11/98, AV.
36 * Fixed a bunch of races (and pair of leaks). Probably not the best way,
37 * but this one obviously doesn't introduce deadlocks. Later. BTW, found
38 * one race (and leak) in BSD implementation.
39 * OK, that's better. ANOTHER race and leak in BSD variant. There always
40 * is one more bug... 10/11/98, AV.
42 * Oh, fsck... Oopsable SMP race in do_process_acct() - we must hold
43 * ->mmap_lock to walk the vma list of current->mm. Nasty, since it leaks
44 * a struct file opened for write. Fixed. 2/6/2000, AV.
48 #include <linux/slab.h>
49 #include <linux/acct.h>
50 #include <linux/capability.h>
51 #include <linux/file.h>
52 #include <linux/tty.h>
53 #include <linux/security.h>
54 #include <linux/vfs.h>
55 #include <linux/jiffies.h>
56 #include <linux/times.h>
57 #include <linux/syscalls.h>
58 #include <linux/mount.h>
59 #include <linux/uaccess.h>
60 #include <linux/sched/cputime.h>
62 #include <asm/div64.h>
63 #include <linux/blkdev.h> /* sector_div */
64 #include <linux/pid_namespace.h>
65 #include <linux/fs_pin.h>
68 * These constants control the amount of freespace that suspend and
69 * resume the process accounting system, and the time delay between
71 * Turned into sysctl-controllable parameters. AV, 12/11/98
74 int acct_parm[3] = {4, 2, 30};
75 #define RESUME (acct_parm[0]) /* >foo% free space - resume */
76 #define SUSPEND (acct_parm[1]) /* <foo% free space - suspend */
77 #define ACCT_TIMEOUT (acct_parm[2]) /* foo second timeout between checks */
80 * External references and all of the globals.
83 struct bsd_acct_struct {
89 unsigned long needcheck;
91 struct pid_namespace *ns;
92 struct work_struct work;
93 struct completion done;
96 static void do_acct_process(struct bsd_acct_struct *acct);
99 * Check the amount of free space and suspend/resume accordingly.
101 static int check_free_space(struct bsd_acct_struct *acct)
105 if (time_is_after_jiffies(acct->needcheck))
109 if (vfs_statfs(&acct->file->f_path, &sbuf))
113 u64 suspend = sbuf.f_blocks * SUSPEND;
114 do_div(suspend, 100);
115 if (sbuf.f_bavail <= suspend) {
117 pr_info("Process accounting paused\n");
120 u64 resume = sbuf.f_blocks * RESUME;
122 if (sbuf.f_bavail >= resume) {
124 pr_info("Process accounting resumed\n");
128 acct->needcheck = jiffies + ACCT_TIMEOUT*HZ;
133 static void acct_put(struct bsd_acct_struct *p)
135 if (atomic_long_dec_and_test(&p->count))
139 static inline struct bsd_acct_struct *to_acct(struct fs_pin *p)
141 return p ? container_of(p, struct bsd_acct_struct, pin) : NULL;
144 static struct bsd_acct_struct *acct_get(struct pid_namespace *ns)
146 struct bsd_acct_struct *res;
150 res = to_acct(READ_ONCE(ns->bacct));
155 if (!atomic_long_inc_not_zero(&res->count)) {
161 mutex_lock(&res->lock);
162 if (res != to_acct(READ_ONCE(ns->bacct))) {
163 mutex_unlock(&res->lock);
170 static void acct_pin_kill(struct fs_pin *pin)
172 struct bsd_acct_struct *acct = to_acct(pin);
173 mutex_lock(&acct->lock);
174 do_acct_process(acct);
175 schedule_work(&acct->work);
176 wait_for_completion(&acct->done);
177 cmpxchg(&acct->ns->bacct, pin, NULL);
178 mutex_unlock(&acct->lock);
183 static void close_work(struct work_struct *work)
185 struct bsd_acct_struct *acct = container_of(work, struct bsd_acct_struct, work);
186 struct file *file = acct->file;
187 if (file->f_op->flush)
188 file->f_op->flush(file, NULL);
190 complete(&acct->done);
193 static int acct_on(struct filename *pathname)
196 struct vfsmount *mnt, *internal;
197 struct pid_namespace *ns = task_active_pid_ns(current);
198 struct bsd_acct_struct *acct;
202 acct = kzalloc(sizeof(struct bsd_acct_struct), GFP_KERNEL);
206 /* Difference from BSD - they don't do O_APPEND */
207 file = file_open_name(pathname, O_WRONLY|O_APPEND|O_LARGEFILE, 0);
210 return PTR_ERR(file);
213 if (!S_ISREG(file_inode(file)->i_mode)) {
215 filp_close(file, NULL);
219 if (!(file->f_mode & FMODE_CAN_WRITE)) {
221 filp_close(file, NULL);
224 internal = mnt_clone_internal(&file->f_path);
225 if (IS_ERR(internal)) {
227 filp_close(file, NULL);
228 return PTR_ERR(internal);
230 err = __mnt_want_write(internal);
234 filp_close(file, NULL);
237 mnt = file->f_path.mnt;
238 file->f_path.mnt = internal;
240 atomic_long_set(&acct->count, 1);
241 init_fs_pin(&acct->pin, acct_pin_kill);
243 acct->needcheck = jiffies;
245 mutex_init(&acct->lock);
246 INIT_WORK(&acct->work, close_work);
247 init_completion(&acct->done);
248 mutex_lock_nested(&acct->lock, 1); /* nobody has seen it yet */
249 pin_insert(&acct->pin, mnt);
252 old = xchg(&ns->bacct, &acct->pin);
253 mutex_unlock(&acct->lock);
255 __mnt_drop_write(mnt);
260 static DEFINE_MUTEX(acct_on_mutex);
263 * sys_acct - enable/disable process accounting
264 * @name: file name for accounting records or NULL to shutdown accounting
266 * sys_acct() is the only system call needed to implement process
267 * accounting. It takes the name of the file where accounting records
268 * should be written. If the filename is NULL, accounting will be
271 * Returns: 0 for success or negative errno values for failure.
273 SYSCALL_DEFINE1(acct, const char __user *, name)
277 if (!capable(CAP_SYS_PACCT))
281 struct filename *tmp = getname(name);
285 mutex_lock(&acct_on_mutex);
286 error = acct_on(tmp);
287 mutex_unlock(&acct_on_mutex);
291 pin_kill(task_active_pid_ns(current)->bacct);
297 void acct_exit_ns(struct pid_namespace *ns)
304 * encode an unsigned long into a comp_t
306 * This routine has been adopted from the encode_comp_t() function in
307 * the kern_acct.c file of the FreeBSD operating system. The encoding
308 * is a 13-bit fraction with a 3-bit (base 8) exponent.
311 #define MANTSIZE 13 /* 13 bit mantissa. */
312 #define EXPSIZE 3 /* Base 8 (3 bit) exponent. */
313 #define MAXFRACT ((1 << MANTSIZE) - 1) /* Maximum fractional value. */
315 static comp_t encode_comp_t(unsigned long value)
320 while (value > MAXFRACT) {
321 rnd = value & (1 << (EXPSIZE - 1)); /* Round up? */
322 value >>= EXPSIZE; /* Base 8 exponent == 3 bit shift. */
327 * If we need to round up, do it (and handle overflow correctly).
329 if (rnd && (++value > MAXFRACT)) {
335 * Clean it up and polish it off.
337 exp <<= MANTSIZE; /* Shift the exponent into place */
338 exp += value; /* and add on the mantissa. */
342 #if ACCT_VERSION == 1 || ACCT_VERSION == 2
344 * encode an u64 into a comp2_t (24 bits)
346 * Format: 5 bit base 2 exponent, 20 bits mantissa.
347 * The leading bit of the mantissa is not stored, but implied for
348 * non-zero exponents.
349 * Largest encodable value is 50 bits.
352 #define MANTSIZE2 20 /* 20 bit mantissa. */
353 #define EXPSIZE2 5 /* 5 bit base 2 exponent. */
354 #define MAXFRACT2 ((1ul << MANTSIZE2) - 1) /* Maximum fractional value. */
355 #define MAXEXP2 ((1 << EXPSIZE2) - 1) /* Maximum exponent. */
357 static comp2_t encode_comp2_t(u64 value)
361 exp = (value > (MAXFRACT2>>1));
363 while (value > MAXFRACT2) {
370 * If we need to round up, do it (and handle overflow correctly).
372 if (rnd && (++value > MAXFRACT2)) {
378 /* Overflow. Return largest representable number instead. */
379 return (1ul << (MANTSIZE2+EXPSIZE2-1)) - 1;
381 return (value & (MAXFRACT2>>1)) | (exp << (MANTSIZE2-1));
384 #elif ACCT_VERSION == 3
386 * encode an u64 into a 32 bit IEEE float
388 static u32 encode_float(u64 value)
395 while ((s64)value > 0) {
399 u = (u32)(value >> 40) & 0x7fffffu;
400 return u | (exp << 23);
405 * Write an accounting entry for an exiting process
407 * The acct_process() call is the workhorse of the process
408 * accounting system. The struct acct is built here and then written
409 * into the accounting file. This function should only be called from
410 * do_exit() or when switching to a different output file.
413 static void fill_ac(acct_t *ac)
415 struct pacct_struct *pacct = ¤t->signal->pacct;
416 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 btime = ktime_get_real_seconds() - elapsed;
451 ac->ac_btime = clamp_t(time64_t, btime, 0, U32_MAX);
456 spin_lock_irq(¤t->sighand->siglock);
457 tty = current->signal->tty; /* Safe as we hold the siglock */
458 ac->ac_tty = tty ? old_encode_dev(tty_devnum(tty)) : 0;
459 ac->ac_utime = encode_comp_t(nsec_to_AHZ(pacct->ac_utime));
460 ac->ac_stime = encode_comp_t(nsec_to_AHZ(pacct->ac_stime));
461 ac->ac_flag = pacct->ac_flag;
462 ac->ac_mem = encode_comp_t(pacct->ac_mem);
463 ac->ac_minflt = encode_comp_t(pacct->ac_minflt);
464 ac->ac_majflt = encode_comp_t(pacct->ac_majflt);
465 ac->ac_exitcode = pacct->ac_exitcode;
466 spin_unlock_irq(¤t->sighand->siglock);
469 * do_acct_process does all actual work. Caller holds the reference to file.
471 static void do_acct_process(struct bsd_acct_struct *acct)
475 const struct cred *orig_cred;
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;
501 #elif ACCT_VERSION == 3
503 struct pid_namespace *ns = acct->ns;
505 ac.ac_pid = task_tgid_nr_ns(current, ns);
507 ac.ac_ppid = task_tgid_nr_ns(rcu_dereference(current->real_parent),
513 * Get freeze protection. If the fs is frozen, just skip the write
514 * as we could deadlock the system otherwise.
516 if (file_start_write_trylock(file)) {
517 /* it's been opened O_APPEND, so position is irrelevant */
519 __kernel_write(file, &ac, sizeof(acct_t), &pos);
520 file_end_write(file);
523 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = flim;
524 revert_creds(orig_cred);
528 * acct_collect - collect accounting information into pacct_struct
529 * @exitcode: task exit code
530 * @group_dead: not 0, if this thread is the last one in the process.
532 void acct_collect(long exitcode, int group_dead)
534 struct pacct_struct *pacct = ¤t->signal->pacct;
536 unsigned long vsize = 0;
538 if (group_dead && current->mm) {
539 struct vm_area_struct *vma;
541 mmap_read_lock(current->mm);
542 vma = current->mm->mmap;
544 vsize += vma->vm_end - vma->vm_start;
547 mmap_read_unlock(current->mm);
550 spin_lock_irq(¤t->sighand->siglock);
552 pacct->ac_mem = vsize / 1024;
553 if (thread_group_leader(current)) {
554 pacct->ac_exitcode = exitcode;
555 if (current->flags & PF_FORKNOEXEC)
556 pacct->ac_flag |= AFORK;
558 if (current->flags & PF_SUPERPRIV)
559 pacct->ac_flag |= ASU;
560 if (current->flags & PF_DUMPCORE)
561 pacct->ac_flag |= ACORE;
562 if (current->flags & PF_SIGNALED)
563 pacct->ac_flag |= AXSIG;
565 task_cputime(current, &utime, &stime);
566 pacct->ac_utime += utime;
567 pacct->ac_stime += stime;
568 pacct->ac_minflt += current->min_flt;
569 pacct->ac_majflt += current->maj_flt;
570 spin_unlock_irq(¤t->sighand->siglock);
573 static void slow_acct_process(struct pid_namespace *ns)
575 for ( ; ns; ns = ns->parent) {
576 struct bsd_acct_struct *acct = acct_get(ns);
578 do_acct_process(acct);
579 mutex_unlock(&acct->lock);
586 * acct_process - handles process accounting for an exiting task
588 void acct_process(void)
590 struct pid_namespace *ns;
593 * This loop is safe lockless, since current is still
594 * alive and holds its namespace, which in turn holds
597 for (ns = task_active_pid_ns(current); ns != NULL; ns = ns->parent) {
602 slow_acct_process(ns);