1/*
2 *  linux/kernel/acct.c
3 *
4 *  BSD Process Accounting for Linux
5 *
6 *  Author: Marco van Wieringen <mvw@planets.elm.net>
7 *
8 *  Some code based on ideas and code from:
9 *  Thomas K. Dyas <tdyas@eden.rutgers.edu>
10 *
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.
16 *
17 * (C) Copyright 1995 - 1997 Marco van Wieringen - ELM Consultancy B.V.
18 *
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.
23 *
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.
26 *
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.
34 *
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.
40 *
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.
44 */
45
46#include <linux/mm.h>
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>
63
64/*
65 * These constants control the amount of freespace that suspend and
66 * resume the process accounting system, and the time delay between
67 * each check.
68 * Turned into sysctl-controllable parameters. AV, 12/11/98
69 */
70
71int 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 */
75
76/*
77 * External references and all of the globals.
78 */
79
80struct bsd_acct_struct {
81	struct fs_pin		pin;
82	atomic_long_t		count;
83	struct rcu_head		rcu;
84	struct mutex		lock;
85	int			active;
86	unsigned long		needcheck;
87	struct file		*file;
88	struct pid_namespace	*ns;
89	struct work_struct	work;
90	struct completion	done;
91};
92
93static void do_acct_process(struct bsd_acct_struct *acct);
94
95/*
96 * Check the amount of free space and suspend/resume accordingly.
97 */
98static int check_free_space(struct bsd_acct_struct *acct)
99{
100	struct kstatfs sbuf;
101
102	if (time_is_before_jiffies(acct->needcheck))
103		goto out;
104
105	/* May block */
106	if (vfs_statfs(&acct->file->f_path, &sbuf))
107		goto out;
108
109	if (acct->active) {
110		u64 suspend = sbuf.f_blocks * SUSPEND;
111		do_div(suspend, 100);
112		if (sbuf.f_bavail <= suspend) {
113			acct->active = 0;
114			pr_info("Process accounting paused\n");
115		}
116	} else {
117		u64 resume = sbuf.f_blocks * RESUME;
118		do_div(resume, 100);
119		if (sbuf.f_bavail >= resume) {
120			acct->active = 1;
121			pr_info("Process accounting resumed\n");
122		}
123	}
124
125	acct->needcheck = jiffies + ACCT_TIMEOUT*HZ;
126out:
127	return acct->active;
128}
129
130static void acct_put(struct bsd_acct_struct *p)
131{
132	if (atomic_long_dec_and_test(&p->count))
133		kfree_rcu(p, rcu);
134}
135
136static inline struct bsd_acct_struct *to_acct(struct fs_pin *p)
137{
138	return p ? container_of(p, struct bsd_acct_struct, pin) : NULL;
139}
140
141static struct bsd_acct_struct *acct_get(struct pid_namespace *ns)
142{
143	struct bsd_acct_struct *res;
144again:
145	smp_rmb();
146	rcu_read_lock();
147	res = to_acct(ACCESS_ONCE(ns->bacct));
148	if (!res) {
149		rcu_read_unlock();
150		return NULL;
151	}
152	if (!atomic_long_inc_not_zero(&res->count)) {
153		rcu_read_unlock();
154		cpu_relax();
155		goto again;
156	}
157	rcu_read_unlock();
158	mutex_lock(&res->lock);
159	if (res != to_acct(ACCESS_ONCE(ns->bacct))) {
160		mutex_unlock(&res->lock);
161		acct_put(res);
162		goto again;
163	}
164	return res;
165}
166
167static void acct_pin_kill(struct fs_pin *pin)
168{
169	struct bsd_acct_struct *acct = to_acct(pin);
170	mutex_lock(&acct->lock);
171	do_acct_process(acct);
172	schedule_work(&acct->work);
173	wait_for_completion(&acct->done);
174	cmpxchg(&acct->ns->bacct, pin, NULL);
175	mutex_unlock(&acct->lock);
176	pin_remove(pin);
177	acct_put(acct);
178}
179
180static void close_work(struct work_struct *work)
181{
182	struct bsd_acct_struct *acct = container_of(work, struct bsd_acct_struct, work);
183	struct file *file = acct->file;
184	if (file->f_op->flush)
185		file->f_op->flush(file, NULL);
186	__fput_sync(file);
187	complete(&acct->done);
188}
189
190static int acct_on(struct filename *pathname)
191{
192	struct file *file;
193	struct vfsmount *mnt, *internal;
194	struct pid_namespace *ns = task_active_pid_ns(current);
195	struct bsd_acct_struct *acct;
196	struct fs_pin *old;
197	int err;
198
199	acct = kzalloc(sizeof(struct bsd_acct_struct), GFP_KERNEL);
200	if (!acct)
201		return -ENOMEM;
202
203	/* Difference from BSD - they don't do O_APPEND */
204	file = file_open_name(pathname, O_WRONLY|O_APPEND|O_LARGEFILE, 0);
205	if (IS_ERR(file)) {
206		kfree(acct);
207		return PTR_ERR(file);
208	}
209
210	if (!S_ISREG(file_inode(file)->i_mode)) {
211		kfree(acct);
212		filp_close(file, NULL);
213		return -EACCES;
214	}
215
216	if (!(file->f_mode & FMODE_CAN_WRITE)) {
217		kfree(acct);
218		filp_close(file, NULL);
219		return -EIO;
220	}
221	internal = mnt_clone_internal(&file->f_path);
222	if (IS_ERR(internal)) {
223		kfree(acct);
224		filp_close(file, NULL);
225		return PTR_ERR(internal);
226	}
227	err = mnt_want_write(internal);
228	if (err) {
229		mntput(internal);
230		kfree(acct);
231		filp_close(file, NULL);
232		return err;
233	}
234	mnt = file->f_path.mnt;
235	file->f_path.mnt = internal;
236
237	atomic_long_set(&acct->count, 1);
238	init_fs_pin(&acct->pin, acct_pin_kill);
239	acct->file = file;
240	acct->needcheck = jiffies;
241	acct->ns = ns;
242	mutex_init(&acct->lock);
243	INIT_WORK(&acct->work, close_work);
244	init_completion(&acct->done);
245	mutex_lock_nested(&acct->lock, 1);	/* nobody has seen it yet */
246	pin_insert(&acct->pin, mnt);
247
248	rcu_read_lock();
249	old = xchg(&ns->bacct, &acct->pin);
250	mutex_unlock(&acct->lock);
251	pin_kill(old);
252	mnt_drop_write(mnt);
253	mntput(mnt);
254	return 0;
255}
256
257static DEFINE_MUTEX(acct_on_mutex);
258
259/**
260 * sys_acct - enable/disable process accounting
261 * @name: file name for accounting records or NULL to shutdown accounting
262 *
263 * Returns 0 for success or negative errno values for failure.
264 *
265 * sys_acct() is the only system call needed to implement process
266 * accounting. It takes the name of the file where accounting records
267 * should be written. If the filename is NULL, accounting will be
268 * shutdown.
269 */
270SYSCALL_DEFINE1(acct, const char __user *, name)
271{
272	int error = 0;
273
274	if (!capable(CAP_SYS_PACCT))
275		return -EPERM;
276
277	if (name) {
278		struct filename *tmp = getname(name);
279
280		if (IS_ERR(tmp))
281			return PTR_ERR(tmp);
282		mutex_lock(&acct_on_mutex);
283		error = acct_on(tmp);
284		mutex_unlock(&acct_on_mutex);
285		putname(tmp);
286	} else {
287		rcu_read_lock();
288		pin_kill(task_active_pid_ns(current)->bacct);
289	}
290
291	return error;
292}
293
294void acct_exit_ns(struct pid_namespace *ns)
295{
296	rcu_read_lock();
297	pin_kill(ns->bacct);
298}
299
300/*
301 *  encode an unsigned long into a comp_t
302 *
303 *  This routine has been adopted from the encode_comp_t() function in
304 *  the kern_acct.c file of the FreeBSD operating system. The encoding
305 *  is a 13-bit fraction with a 3-bit (base 8) exponent.
306 */
307
308#define	MANTSIZE	13			/* 13 bit mantissa. */
309#define	EXPSIZE		3			/* Base 8 (3 bit) exponent. */
310#define	MAXFRACT	((1 << MANTSIZE) - 1)	/* Maximum fractional value. */
311
312static comp_t encode_comp_t(unsigned long value)
313{
314	int exp, rnd;
315
316	exp = rnd = 0;
317	while (value > MAXFRACT) {
318		rnd = value & (1 << (EXPSIZE - 1));	/* Round up? */
319		value >>= EXPSIZE;	/* Base 8 exponent == 3 bit shift. */
320		exp++;
321	}
322
323	/*
324	 * If we need to round up, do it (and handle overflow correctly).
325	 */
326	if (rnd && (++value > MAXFRACT)) {
327		value >>= EXPSIZE;
328		exp++;
329	}
330
331	/*
332	 * Clean it up and polish it off.
333	 */
334	exp <<= MANTSIZE;		/* Shift the exponent into place */
335	exp += value;			/* and add on the mantissa. */
336	return exp;
337}
338
339#if ACCT_VERSION == 1 || ACCT_VERSION == 2
340/*
341 * encode an u64 into a comp2_t (24 bits)
342 *
343 * Format: 5 bit base 2 exponent, 20 bits mantissa.
344 * The leading bit of the mantissa is not stored, but implied for
345 * non-zero exponents.
346 * Largest encodable value is 50 bits.
347 */
348
349#define MANTSIZE2       20                      /* 20 bit mantissa. */
350#define EXPSIZE2        5                       /* 5 bit base 2 exponent. */
351#define MAXFRACT2       ((1ul << MANTSIZE2) - 1) /* Maximum fractional value. */
352#define MAXEXP2         ((1 << EXPSIZE2) - 1)    /* Maximum exponent. */
353
354static comp2_t encode_comp2_t(u64 value)
355{
356	int exp, rnd;
357
358	exp = (value > (MAXFRACT2>>1));
359	rnd = 0;
360	while (value > MAXFRACT2) {
361		rnd = value & 1;
362		value >>= 1;
363		exp++;
364	}
365
366	/*
367	 * If we need to round up, do it (and handle overflow correctly).
368	 */
369	if (rnd && (++value > MAXFRACT2)) {
370		value >>= 1;
371		exp++;
372	}
373
374	if (exp > MAXEXP2) {
375		/* Overflow. Return largest representable number instead. */
376		return (1ul << (MANTSIZE2+EXPSIZE2-1)) - 1;
377	} else {
378		return (value & (MAXFRACT2>>1)) | (exp << (MANTSIZE2-1));
379	}
380}
381#endif
382
383#if ACCT_VERSION == 3
384/*
385 * encode an u64 into a 32 bit IEEE float
386 */
387static u32 encode_float(u64 value)
388{
389	unsigned exp = 190;
390	unsigned u;
391
392	if (value == 0)
393		return 0;
394	while ((s64)value > 0) {
395		value <<= 1;
396		exp--;
397	}
398	u = (u32)(value >> 40) & 0x7fffffu;
399	return u | (exp << 23);
400}
401#endif
402
403/*
404 *  Write an accounting entry for an exiting process
405 *
406 *  The acct_process() call is the workhorse of the process
407 *  accounting system. The struct acct is built here and then written
408 *  into the accounting file. This function should only be called from
409 *  do_exit() or when switching to a different output file.
410 */
411
412static void fill_ac(acct_t *ac)
413{
414	struct pacct_struct *pacct = &current->signal->pacct;
415	u64 elapsed, run_time;
416	struct tty_struct *tty;
417
418	/*
419	 * Fill the accounting struct with the needed info as recorded
420	 * by the different kernel functions.
421	 */
422	memset(ac, 0, sizeof(acct_t));
423
424	ac->ac_version = ACCT_VERSION | ACCT_BYTEORDER;
425	strlcpy(ac->ac_comm, current->comm, sizeof(ac->ac_comm));
426
427	/* calculate run_time in nsec*/
428	run_time = ktime_get_ns();
429	run_time -= current->group_leader->start_time;
430	/* convert nsec -> AHZ */
431	elapsed = nsec_to_AHZ(run_time);
432#if ACCT_VERSION == 3
433	ac->ac_etime = encode_float(elapsed);
434#else
435	ac->ac_etime = encode_comp_t(elapsed < (unsigned long) -1l ?
436				(unsigned long) elapsed : (unsigned long) -1l);
437#endif
438#if ACCT_VERSION == 1 || ACCT_VERSION == 2
439	{
440		/* new enlarged etime field */
441		comp2_t etime = encode_comp2_t(elapsed);
442
443		ac->ac_etime_hi = etime >> 16;
444		ac->ac_etime_lo = (u16) etime;
445	}
446#endif
447	do_div(elapsed, AHZ);
448	ac->ac_btime = get_seconds() - elapsed;
449#if ACCT_VERSION==2
450	ac->ac_ahz = AHZ;
451#endif
452
453	spin_lock_irq(&current->sighand->siglock);
454	tty = current->signal->tty;	/* Safe as we hold the siglock */
455	ac->ac_tty = tty ? old_encode_dev(tty_devnum(tty)) : 0;
456	ac->ac_utime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_utime)));
457	ac->ac_stime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_stime)));
458	ac->ac_flag = pacct->ac_flag;
459	ac->ac_mem = encode_comp_t(pacct->ac_mem);
460	ac->ac_minflt = encode_comp_t(pacct->ac_minflt);
461	ac->ac_majflt = encode_comp_t(pacct->ac_majflt);
462	ac->ac_exitcode = pacct->ac_exitcode;
463	spin_unlock_irq(&current->sighand->siglock);
464}
465/*
466 *  do_acct_process does all actual work. Caller holds the reference to file.
467 */
468static void do_acct_process(struct bsd_acct_struct *acct)
469{
470	acct_t ac;
471	unsigned long flim;
472	const struct cred *orig_cred;
473	struct file *file = acct->file;
474
475	/*
476	 * Accounting records are not subject to resource limits.
477	 */
478	flim = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
479	current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
480	/* Perform file operations on behalf of whoever enabled accounting */
481	orig_cred = override_creds(file->f_cred);
482
483	/*
484	 * First check to see if there is enough free_space to continue
485	 * the process accounting system.
486	 */
487	if (!check_free_space(acct))
488		goto out;
489
490	fill_ac(&ac);
491	/* we really need to bite the bullet and change layout */
492	ac.ac_uid = from_kuid_munged(file->f_cred->user_ns, orig_cred->uid);
493	ac.ac_gid = from_kgid_munged(file->f_cred->user_ns, orig_cred->gid);
494#if ACCT_VERSION == 1 || ACCT_VERSION == 2
495	/* backward-compatible 16 bit fields */
496	ac.ac_uid16 = ac.ac_uid;
497	ac.ac_gid16 = ac.ac_gid;
498#endif
499#if ACCT_VERSION == 3
500	{
501		struct pid_namespace *ns = acct->ns;
502
503		ac.ac_pid = task_tgid_nr_ns(current, ns);
504		rcu_read_lock();
505		ac.ac_ppid = task_tgid_nr_ns(rcu_dereference(current->real_parent),
506					     ns);
507		rcu_read_unlock();
508	}
509#endif
510	/*
511	 * Get freeze protection. If the fs is frozen, just skip the write
512	 * as we could deadlock the system otherwise.
513	 */
514	if (file_start_write_trylock(file)) {
515		/* it's been opened O_APPEND, so position is irrelevant */
516		loff_t pos = 0;
517		__kernel_write(file, (char *)&ac, sizeof(acct_t), &pos);
518		file_end_write(file);
519	}
520out:
521	current->signal->rlim[RLIMIT_FSIZE].rlim_cur = flim;
522	revert_creds(orig_cred);
523}
524
525/**
526 * acct_collect - collect accounting information into pacct_struct
527 * @exitcode: task exit code
528 * @group_dead: not 0, if this thread is the last one in the process.
529 */
530void acct_collect(long exitcode, int group_dead)
531{
532	struct pacct_struct *pacct = &current->signal->pacct;
533	cputime_t utime, stime;
534	unsigned long vsize = 0;
535
536	if (group_dead && current->mm) {
537		struct vm_area_struct *vma;
538
539		down_read(&current->mm->mmap_sem);
540		vma = current->mm->mmap;
541		while (vma) {
542			vsize += vma->vm_end - vma->vm_start;
543			vma = vma->vm_next;
544		}
545		up_read(&current->mm->mmap_sem);
546	}
547
548	spin_lock_irq(&current->sighand->siglock);
549	if (group_dead)
550		pacct->ac_mem = vsize / 1024;
551	if (thread_group_leader(current)) {
552		pacct->ac_exitcode = exitcode;
553		if (current->flags & PF_FORKNOEXEC)
554			pacct->ac_flag |= AFORK;
555	}
556	if (current->flags & PF_SUPERPRIV)
557		pacct->ac_flag |= ASU;
558	if (current->flags & PF_DUMPCORE)
559		pacct->ac_flag |= ACORE;
560	if (current->flags & PF_SIGNALED)
561		pacct->ac_flag |= AXSIG;
562	task_cputime(current, &utime, &stime);
563	pacct->ac_utime += utime;
564	pacct->ac_stime += stime;
565	pacct->ac_minflt += current->min_flt;
566	pacct->ac_majflt += current->maj_flt;
567	spin_unlock_irq(&current->sighand->siglock);
568}
569
570static void slow_acct_process(struct pid_namespace *ns)
571{
572	for ( ; ns; ns = ns->parent) {
573		struct bsd_acct_struct *acct = acct_get(ns);
574		if (acct) {
575			do_acct_process(acct);
576			mutex_unlock(&acct->lock);
577			acct_put(acct);
578		}
579	}
580}
581
582/**
583 * acct_process
584 *
585 * handles process accounting for an exiting task
586 */
587void acct_process(void)
588{
589	struct pid_namespace *ns;
590
591	/*
592	 * This loop is safe lockless, since current is still
593	 * alive and holds its namespace, which in turn holds
594	 * its parent.
595	 */
596	for (ns = task_active_pid_ns(current); ns != NULL; ns = ns->parent) {
597		if (ns->bacct)
598			break;
599	}
600	if (unlikely(ns))
601		slow_acct_process(ns);
602}
603