• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 
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 */
75 
76 /*
77  * External references and all of the globals.
78  */
79 static void do_acct_process(struct bsd_acct_struct *acct);
80 
81 struct bsd_acct_struct {
82 	struct fs_pin		pin;
83 	struct mutex		lock;
84 	int			active;
85 	unsigned long		needcheck;
86 	struct file		*file;
87 	struct pid_namespace	*ns;
88 	struct work_struct	work;
89 	struct completion	done;
90 };
91 
92 /*
93  * Check the amount of free space and suspend/resume accordingly.
94  */
check_free_space(struct bsd_acct_struct * acct)95 static int check_free_space(struct bsd_acct_struct *acct)
96 {
97 	struct kstatfs sbuf;
98 
99 	if (time_is_after_jiffies(acct->needcheck))
100 		goto out;
101 
102 	/* May block */
103 	if (vfs_statfs(&acct->file->f_path, &sbuf))
104 		goto out;
105 
106 	if (acct->active) {
107 		u64 suspend = sbuf.f_blocks * SUSPEND;
108 		do_div(suspend, 100);
109 		if (sbuf.f_bavail <= suspend) {
110 			acct->active = 0;
111 			pr_info("Process accounting paused\n");
112 		}
113 	} else {
114 		u64 resume = sbuf.f_blocks * RESUME;
115 		do_div(resume, 100);
116 		if (sbuf.f_bavail >= resume) {
117 			acct->active = 1;
118 			pr_info("Process accounting resumed\n");
119 		}
120 	}
121 
122 	acct->needcheck = jiffies + ACCT_TIMEOUT*HZ;
123 out:
124 	return acct->active;
125 }
126 
acct_get(struct pid_namespace * ns)127 static struct bsd_acct_struct *acct_get(struct pid_namespace *ns)
128 {
129 	struct bsd_acct_struct *res;
130 again:
131 	smp_rmb();
132 	rcu_read_lock();
133 	res = ACCESS_ONCE(ns->bacct);
134 	if (!res) {
135 		rcu_read_unlock();
136 		return NULL;
137 	}
138 	if (!atomic_long_inc_not_zero(&res->pin.count)) {
139 		rcu_read_unlock();
140 		cpu_relax();
141 		goto again;
142 	}
143 	rcu_read_unlock();
144 	mutex_lock(&res->lock);
145 	if (!res->ns) {
146 		mutex_unlock(&res->lock);
147 		pin_put(&res->pin);
148 		goto again;
149 	}
150 	return res;
151 }
152 
close_work(struct work_struct * work)153 static void close_work(struct work_struct *work)
154 {
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);
159 	__fput_sync(file);
160 	complete(&acct->done);
161 }
162 
acct_kill(struct bsd_acct_struct * acct,struct bsd_acct_struct * new)163 static void acct_kill(struct bsd_acct_struct *acct,
164 		      struct bsd_acct_struct *new)
165 {
166 	if (acct) {
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);
174 		ns->bacct = new;
175 		acct->ns = NULL;
176 		atomic_long_dec(&acct->pin.count);
177 		mutex_unlock(&acct->lock);
178 		pin_put(&acct->pin);
179 	}
180 }
181 
acct_pin_kill(struct fs_pin * pin)182 static void acct_pin_kill(struct fs_pin *pin)
183 {
184 	struct bsd_acct_struct *acct;
185 	acct = container_of(pin, struct bsd_acct_struct, pin);
186 	mutex_lock(&acct->lock);
187 	if (!acct->ns) {
188 		mutex_unlock(&acct->lock);
189 		pin_put(pin);
190 		acct = NULL;
191 	}
192 	acct_kill(acct, NULL);
193 }
194 
acct_on(struct filename * pathname)195 static int acct_on(struct filename *pathname)
196 {
197 	struct file *file;
198 	struct vfsmount *mnt, *internal;
199 	struct pid_namespace *ns = task_active_pid_ns(current);
200 	struct bsd_acct_struct *acct, *old;
201 	int err;
202 
203 	acct = kzalloc(sizeof(struct bsd_acct_struct), GFP_KERNEL);
204 	if (!acct)
205 		return -ENOMEM;
206 
207 	/* Difference from BSD - they don't do O_APPEND */
208 	file = file_open_name(pathname, O_WRONLY|O_APPEND|O_LARGEFILE, 0);
209 	if (IS_ERR(file)) {
210 		kfree(acct);
211 		return PTR_ERR(file);
212 	}
213 
214 	if (!S_ISREG(file_inode(file)->i_mode)) {
215 		kfree(acct);
216 		filp_close(file, NULL);
217 		return -EACCES;
218 	}
219 
220 	if (!file->f_op->write) {
221 		kfree(acct);
222 		filp_close(file, NULL);
223 		return -EIO;
224 	}
225 	internal = mnt_clone_internal(&file->f_path);
226 	if (IS_ERR(internal)) {
227 		kfree(acct);
228 		filp_close(file, NULL);
229 		return PTR_ERR(internal);
230 	}
231 	err = mnt_want_write(internal);
232 	if (err) {
233 		mntput(internal);
234 		kfree(acct);
235 		filp_close(file, NULL);
236 		return err;
237 	}
238 	mnt = file->f_path.mnt;
239 	file->f_path.mnt = internal;
240 
241 	atomic_long_set(&acct->pin.count, 1);
242 	acct->pin.kill = acct_pin_kill;
243 	acct->file = file;
244 	acct->needcheck = jiffies;
245 	acct->ns = ns;
246 	mutex_init(&acct->lock);
247 	mutex_lock_nested(&acct->lock, 1);	/* nobody has seen it yet */
248 	pin_insert(&acct->pin, mnt);
249 
250 	old = acct_get(ns);
251 	if (old)
252 		acct_kill(old, acct);
253 	else
254 		ns->bacct = acct;
255 	mutex_unlock(&acct->lock);
256 	mnt_drop_write(mnt);
257 	mntput(mnt);
258 	return 0;
259 }
260 
261 static DEFINE_MUTEX(acct_on_mutex);
262 
263 /**
264  * sys_acct - enable/disable process accounting
265  * @name: file name for accounting records or NULL to shutdown accounting
266  *
267  * Returns 0 for success or negative errno values for failure.
268  *
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
272  * shutdown.
273  */
SYSCALL_DEFINE1(acct,const char __user *,name)274 SYSCALL_DEFINE1(acct, const char __user *, name)
275 {
276 	int error = 0;
277 
278 	if (!capable(CAP_SYS_PACCT))
279 		return -EPERM;
280 
281 	if (name) {
282 		struct filename *tmp = getname(name);
283 
284 		if (IS_ERR(tmp))
285 			return PTR_ERR(tmp);
286 		mutex_lock(&acct_on_mutex);
287 		error = acct_on(tmp);
288 		mutex_unlock(&acct_on_mutex);
289 		putname(tmp);
290 	} else {
291 		acct_kill(acct_get(task_active_pid_ns(current)), NULL);
292 	}
293 
294 	return error;
295 }
296 
acct_exit_ns(struct pid_namespace * ns)297 void acct_exit_ns(struct pid_namespace *ns)
298 {
299 	acct_kill(acct_get(ns), NULL);
300 }
301 
302 /*
303  *  encode an unsigned long into a comp_t
304  *
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.
308  */
309 
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. */
313 
encode_comp_t(unsigned long value)314 static comp_t encode_comp_t(unsigned long value)
315 {
316 	int exp, rnd;
317 
318 	exp = rnd = 0;
319 	while (value > MAXFRACT) {
320 		rnd = value & (1 << (EXPSIZE - 1));	/* Round up? */
321 		value >>= EXPSIZE;	/* Base 8 exponent == 3 bit shift. */
322 		exp++;
323 	}
324 
325 	/*
326 	 * If we need to round up, do it (and handle overflow correctly).
327 	 */
328 	if (rnd && (++value > MAXFRACT)) {
329 		value >>= EXPSIZE;
330 		exp++;
331 	}
332 
333 	/*
334 	 * Clean it up and polish it off.
335 	 */
336 	exp <<= MANTSIZE;		/* Shift the exponent into place */
337 	exp += value;			/* and add on the mantissa. */
338 	return exp;
339 }
340 
341 #if ACCT_VERSION == 1 || ACCT_VERSION == 2
342 /*
343  * encode an u64 into a comp2_t (24 bits)
344  *
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.
349  */
350 
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. */
355 
encode_comp2_t(u64 value)356 static comp2_t encode_comp2_t(u64 value)
357 {
358 	int exp, rnd;
359 
360 	exp = (value > (MAXFRACT2>>1));
361 	rnd = 0;
362 	while (value > MAXFRACT2) {
363 		rnd = value & 1;
364 		value >>= 1;
365 		exp++;
366 	}
367 
368 	/*
369 	 * If we need to round up, do it (and handle overflow correctly).
370 	 */
371 	if (rnd && (++value > MAXFRACT2)) {
372 		value >>= 1;
373 		exp++;
374 	}
375 
376 	if (exp > MAXEXP2) {
377 		/* Overflow. Return largest representable number instead. */
378 		return (1ul << (MANTSIZE2+EXPSIZE2-1)) - 1;
379 	} else {
380 		return (value & (MAXFRACT2>>1)) | (exp << (MANTSIZE2-1));
381 	}
382 }
383 #endif
384 
385 #if ACCT_VERSION == 3
386 /*
387  * encode an u64 into a 32 bit IEEE float
388  */
encode_float(u64 value)389 static u32 encode_float(u64 value)
390 {
391 	unsigned exp = 190;
392 	unsigned u;
393 
394 	if (value == 0)
395 		return 0;
396 	while ((s64)value > 0) {
397 		value <<= 1;
398 		exp--;
399 	}
400 	u = (u32)(value >> 40) & 0x7fffffu;
401 	return u | (exp << 23);
402 }
403 #endif
404 
405 /*
406  *  Write an accounting entry for an exiting process
407  *
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.
412  */
413 
fill_ac(acct_t * ac)414 static void fill_ac(acct_t *ac)
415 {
416 	struct pacct_struct *pacct = &current->signal->pacct;
417 	u64 elapsed, run_time;
418 	struct tty_struct *tty;
419 
420 	/*
421 	 * Fill the accounting struct with the needed info as recorded
422 	 * by the different kernel functions.
423 	 */
424 	memset(ac, 0, sizeof(acct_t));
425 
426 	ac->ac_version = ACCT_VERSION | ACCT_BYTEORDER;
427 	strlcpy(ac->ac_comm, current->comm, sizeof(ac->ac_comm));
428 
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);
436 #else
437 	ac->ac_etime = encode_comp_t(elapsed < (unsigned long) -1l ?
438 				(unsigned long) elapsed : (unsigned long) -1l);
439 #endif
440 #if ACCT_VERSION == 1 || ACCT_VERSION == 2
441 	{
442 		/* new enlarged etime field */
443 		comp2_t etime = encode_comp2_t(elapsed);
444 
445 		ac->ac_etime_hi = etime >> 16;
446 		ac->ac_etime_lo = (u16) etime;
447 	}
448 #endif
449 	do_div(elapsed, AHZ);
450 	ac->ac_btime = get_seconds() - elapsed;
451 #if ACCT_VERSION==2
452 	ac->ac_ahz = AHZ;
453 #endif
454 
455 	spin_lock_irq(&current->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(&current->sighand->siglock);
466 }
467 /*
468  *  do_acct_process does all actual work. Caller holds the reference to file.
469  */
do_acct_process(struct bsd_acct_struct * acct)470 static void do_acct_process(struct bsd_acct_struct *acct)
471 {
472 	acct_t ac;
473 	unsigned long flim;
474 	const struct cred *orig_cred;
475 	struct file *file = acct->file;
476 
477 	/*
478 	 * Accounting records are not subject to resource limits.
479 	 */
480 	flim = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
481 	current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
482 	/* Perform file operations on behalf of whoever enabled accounting */
483 	orig_cred = override_creds(file->f_cred);
484 
485 	/*
486 	 * First check to see if there is enough free_space to continue
487 	 * the process accounting system.
488 	 */
489 	if (!check_free_space(acct))
490 		goto out;
491 
492 	fill_ac(&ac);
493 	/* we really need to bite the bullet and change layout */
494 	ac.ac_uid = from_kuid_munged(file->f_cred->user_ns, orig_cred->uid);
495 	ac.ac_gid = from_kgid_munged(file->f_cred->user_ns, orig_cred->gid);
496 #if ACCT_VERSION == 1 || ACCT_VERSION == 2
497 	/* backward-compatible 16 bit fields */
498 	ac.ac_uid16 = ac.ac_uid;
499 	ac.ac_gid16 = ac.ac_gid;
500 #endif
501 #if ACCT_VERSION == 3
502 	{
503 		struct pid_namespace *ns = acct->ns;
504 
505 		ac.ac_pid = task_tgid_nr_ns(current, ns);
506 		rcu_read_lock();
507 		ac.ac_ppid = task_tgid_nr_ns(rcu_dereference(current->real_parent),
508 					     ns);
509 		rcu_read_unlock();
510 	}
511 #endif
512 	/*
513 	 * Get freeze protection. If the fs is frozen, just skip the write
514 	 * as we could deadlock the system otherwise.
515 	 */
516 	if (file_start_write_trylock(file)) {
517 		/* it's been opened O_APPEND, so position is irrelevant */
518 		loff_t pos = 0;
519 		__kernel_write(file, (char *)&ac, sizeof(acct_t), &pos);
520 		file_end_write(file);
521 	}
522 out:
523 	current->signal->rlim[RLIMIT_FSIZE].rlim_cur = flim;
524 	revert_creds(orig_cred);
525 }
526 
527 /**
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.
531  */
acct_collect(long exitcode,int group_dead)532 void acct_collect(long exitcode, int group_dead)
533 {
534 	struct pacct_struct *pacct = &current->signal->pacct;
535 	cputime_t utime, stime;
536 	unsigned long vsize = 0;
537 
538 	if (group_dead && current->mm) {
539 		struct vm_area_struct *vma;
540 
541 		down_read(&current->mm->mmap_sem);
542 		vma = current->mm->mmap;
543 		while (vma) {
544 			vsize += vma->vm_end - vma->vm_start;
545 			vma = vma->vm_next;
546 		}
547 		up_read(&current->mm->mmap_sem);
548 	}
549 
550 	spin_lock_irq(&current->sighand->siglock);
551 	if (group_dead)
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;
557 	}
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;
564 	task_cputime(current, &utime, &stime);
565 	pacct->ac_utime += utime;
566 	pacct->ac_stime += stime;
567 	pacct->ac_minflt += current->min_flt;
568 	pacct->ac_majflt += current->maj_flt;
569 	spin_unlock_irq(&current->sighand->siglock);
570 }
571 
slow_acct_process(struct pid_namespace * ns)572 static void slow_acct_process(struct pid_namespace *ns)
573 {
574 	for ( ; ns; ns = ns->parent) {
575 		struct bsd_acct_struct *acct = acct_get(ns);
576 		if (acct) {
577 			do_acct_process(acct);
578 			mutex_unlock(&acct->lock);
579 			pin_put(&acct->pin);
580 		}
581 	}
582 }
583 
584 /**
585  * acct_process
586  *
587  * handles process accounting for an exiting task
588  */
acct_process(void)589 void acct_process(void)
590 {
591 	struct pid_namespace *ns;
592 
593 	/*
594 	 * This loop is safe lockless, since current is still
595 	 * alive and holds its namespace, which in turn holds
596 	 * its parent.
597 	 */
598 	for (ns = task_active_pid_ns(current); ns != NULL; ns = ns->parent) {
599 		if (ns->bacct)
600 			break;
601 	}
602 	if (unlikely(ns))
603 		slow_acct_process(ns);
604 }
605