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 <asm/uaccess.h>
59 #include <asm/div64.h>
60 #include <linux/blkdev.h> /* sector_div */
61 #include <linux/pid_namespace.h>
62
63 /*
64 * These constants control the amount of freespace that suspend and
65 * resume the process accounting system, and the time delay between
66 * each check.
67 * Turned into sysctl-controllable parameters. AV, 12/11/98
68 */
69
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 */
74
75 /*
76 * External references and all of the globals.
77 */
78 static void do_acct_process(struct bsd_acct_struct *acct,
79 struct pid_namespace *ns, struct file *);
80
81 /*
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.
85 */
86 struct bsd_acct_struct {
87 volatile int active;
88 volatile int needcheck;
89 struct file *file;
90 struct pid_namespace *ns;
91 struct timer_list timer;
92 struct list_head list;
93 };
94
95 static DEFINE_SPINLOCK(acct_lock);
96 static LIST_HEAD(acct_list);
97
98 /*
99 * Called whenever the timer says to check the free space.
100 */
acct_timeout(unsigned long x)101 static void acct_timeout(unsigned long x)
102 {
103 struct bsd_acct_struct *acct = (struct bsd_acct_struct *)x;
104 acct->needcheck = 1;
105 }
106
107 /*
108 * Check the amount of free space and suspend/resume accordingly.
109 */
check_free_space(struct bsd_acct_struct * acct,struct file * file)110 static int check_free_space(struct bsd_acct_struct *acct, struct file *file)
111 {
112 struct kstatfs sbuf;
113 int res;
114 int act;
115 sector_t resume;
116 sector_t suspend;
117
118 spin_lock(&acct_lock);
119 res = acct->active;
120 if (!file || !acct->needcheck)
121 goto out;
122 spin_unlock(&acct_lock);
123
124 /* May block */
125 if (vfs_statfs(file->f_path.dentry, &sbuf))
126 return res;
127 suspend = sbuf.f_blocks * SUSPEND;
128 resume = sbuf.f_blocks * RESUME;
129
130 sector_div(suspend, 100);
131 sector_div(resume, 100);
132
133 if (sbuf.f_bavail <= suspend)
134 act = -1;
135 else if (sbuf.f_bavail >= resume)
136 act = 1;
137 else
138 act = 0;
139
140 /*
141 * If some joker switched acct->file under us we'ld better be
142 * silent and _not_ touch anything.
143 */
144 spin_lock(&acct_lock);
145 if (file != acct->file) {
146 if (act)
147 res = act>0;
148 goto out;
149 }
150
151 if (acct->active) {
152 if (act < 0) {
153 acct->active = 0;
154 printk(KERN_INFO "Process accounting paused\n");
155 }
156 } else {
157 if (act > 0) {
158 acct->active = 1;
159 printk(KERN_INFO "Process accounting resumed\n");
160 }
161 }
162
163 del_timer(&acct->timer);
164 acct->needcheck = 0;
165 acct->timer.expires = jiffies + ACCT_TIMEOUT*HZ;
166 add_timer(&acct->timer);
167 res = acct->active;
168 out:
169 spin_unlock(&acct_lock);
170 return res;
171 }
172
173 /*
174 * Close the old accounting file (if currently open) and then replace
175 * it with file (if non-NULL).
176 *
177 * NOTE: acct_lock MUST be held on entry and exit.
178 */
acct_file_reopen(struct bsd_acct_struct * acct,struct file * file,struct pid_namespace * ns)179 static void acct_file_reopen(struct bsd_acct_struct *acct, struct file *file,
180 struct pid_namespace *ns)
181 {
182 struct file *old_acct = NULL;
183 struct pid_namespace *old_ns = NULL;
184
185 if (acct->file) {
186 old_acct = acct->file;
187 old_ns = acct->ns;
188 del_timer(&acct->timer);
189 acct->active = 0;
190 acct->needcheck = 0;
191 acct->file = NULL;
192 acct->ns = NULL;
193 list_del(&acct->list);
194 }
195 if (file) {
196 acct->file = file;
197 acct->ns = ns;
198 acct->needcheck = 0;
199 acct->active = 1;
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);
205 }
206 if (old_acct) {
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);
212 }
213 }
214
acct_on(char * name)215 static int acct_on(char *name)
216 {
217 struct file *file;
218 int error;
219 struct pid_namespace *ns;
220 struct bsd_acct_struct *acct = NULL;
221
222 /* Difference from BSD - they don't do O_APPEND */
223 file = filp_open(name, O_WRONLY|O_APPEND|O_LARGEFILE, 0);
224 if (IS_ERR(file))
225 return PTR_ERR(file);
226
227 if (!S_ISREG(file->f_path.dentry->d_inode->i_mode)) {
228 filp_close(file, NULL);
229 return -EACCES;
230 }
231
232 if (!file->f_op->write) {
233 filp_close(file, NULL);
234 return -EIO;
235 }
236
237 ns = task_active_pid_ns(current);
238 if (ns->bacct == NULL) {
239 acct = kzalloc(sizeof(struct bsd_acct_struct), GFP_KERNEL);
240 if (acct == NULL) {
241 filp_close(file, NULL);
242 return -ENOMEM;
243 }
244 }
245
246 error = security_acct(file);
247 if (error) {
248 kfree(acct);
249 filp_close(file, NULL);
250 return error;
251 }
252
253 spin_lock(&acct_lock);
254 if (ns->bacct == NULL) {
255 ns->bacct = acct;
256 acct = NULL;
257 }
258
259 mnt_pin(file->f_path.mnt);
260 acct_file_reopen(ns->bacct, file, ns);
261 spin_unlock(&acct_lock);
262
263 mntput(file->f_path.mnt); /* it's pinned, now give up active reference */
264 kfree(acct);
265
266 return 0;
267 }
268
269 /**
270 * sys_acct - enable/disable process accounting
271 * @name: file name for accounting records or NULL to shutdown accounting
272 *
273 * Returns 0 for success or negative errno values for failure.
274 *
275 * sys_acct() is the only system call needed to implement process
276 * accounting. It takes the name of the file where accounting records
277 * should be written. If the filename is NULL, accounting will be
278 * shutdown.
279 */
SYSCALL_DEFINE1(acct,const char __user *,name)280 SYSCALL_DEFINE1(acct, const char __user *, name)
281 {
282 int error;
283
284 if (!capable(CAP_SYS_PACCT))
285 return -EPERM;
286
287 if (name) {
288 char *tmp = getname(name);
289 if (IS_ERR(tmp))
290 return (PTR_ERR(tmp));
291 error = acct_on(tmp);
292 putname(tmp);
293 } else {
294 struct bsd_acct_struct *acct;
295
296 acct = task_active_pid_ns(current)->bacct;
297 if (acct == NULL)
298 return 0;
299
300 error = security_acct(NULL);
301 if (!error) {
302 spin_lock(&acct_lock);
303 acct_file_reopen(acct, NULL, NULL);
304 spin_unlock(&acct_lock);
305 }
306 }
307 return error;
308 }
309
310 /**
311 * acct_auto_close - turn off a filesystem's accounting if it is on
312 * @m: vfsmount being shut down
313 *
314 * If the accounting is turned on for a file in the subtree pointed to
315 * to by m, turn accounting off. Done when m is about to die.
316 */
acct_auto_close_mnt(struct vfsmount * m)317 void acct_auto_close_mnt(struct vfsmount *m)
318 {
319 struct bsd_acct_struct *acct;
320
321 spin_lock(&acct_lock);
322 restart:
323 list_for_each_entry(acct, &acct_list, list)
324 if (acct->file && acct->file->f_path.mnt == m) {
325 acct_file_reopen(acct, NULL, NULL);
326 goto restart;
327 }
328 spin_unlock(&acct_lock);
329 }
330
331 /**
332 * acct_auto_close - turn off a filesystem's accounting if it is on
333 * @sb: super block for the filesystem
334 *
335 * If the accounting is turned on for a file in the filesystem pointed
336 * to by sb, turn accounting off.
337 */
acct_auto_close(struct super_block * sb)338 void acct_auto_close(struct super_block *sb)
339 {
340 struct bsd_acct_struct *acct;
341
342 spin_lock(&acct_lock);
343 restart:
344 list_for_each_entry(acct, &acct_list, list)
345 if (acct->file && acct->file->f_path.mnt->mnt_sb == sb) {
346 acct_file_reopen(acct, NULL, NULL);
347 goto restart;
348 }
349 spin_unlock(&acct_lock);
350 }
351
acct_exit_ns(struct pid_namespace * ns)352 void acct_exit_ns(struct pid_namespace *ns)
353 {
354 struct bsd_acct_struct *acct;
355
356 spin_lock(&acct_lock);
357 acct = ns->bacct;
358 if (acct != NULL) {
359 if (acct->file != NULL)
360 acct_file_reopen(acct, NULL, NULL);
361
362 kfree(acct);
363 }
364 spin_unlock(&acct_lock);
365 }
366
367 /*
368 * encode an unsigned long into a comp_t
369 *
370 * This routine has been adopted from the encode_comp_t() function in
371 * the kern_acct.c file of the FreeBSD operating system. The encoding
372 * is a 13-bit fraction with a 3-bit (base 8) exponent.
373 */
374
375 #define MANTSIZE 13 /* 13 bit mantissa. */
376 #define EXPSIZE 3 /* Base 8 (3 bit) exponent. */
377 #define MAXFRACT ((1 << MANTSIZE) - 1) /* Maximum fractional value. */
378
encode_comp_t(unsigned long value)379 static comp_t encode_comp_t(unsigned long value)
380 {
381 int exp, rnd;
382
383 exp = rnd = 0;
384 while (value > MAXFRACT) {
385 rnd = value & (1 << (EXPSIZE - 1)); /* Round up? */
386 value >>= EXPSIZE; /* Base 8 exponent == 3 bit shift. */
387 exp++;
388 }
389
390 /*
391 * If we need to round up, do it (and handle overflow correctly).
392 */
393 if (rnd && (++value > MAXFRACT)) {
394 value >>= EXPSIZE;
395 exp++;
396 }
397
398 /*
399 * Clean it up and polish it off.
400 */
401 exp <<= MANTSIZE; /* Shift the exponent into place */
402 exp += value; /* and add on the mantissa. */
403 return exp;
404 }
405
406 #if ACCT_VERSION==1 || ACCT_VERSION==2
407 /*
408 * encode an u64 into a comp2_t (24 bits)
409 *
410 * Format: 5 bit base 2 exponent, 20 bits mantissa.
411 * The leading bit of the mantissa is not stored, but implied for
412 * non-zero exponents.
413 * Largest encodable value is 50 bits.
414 */
415
416 #define MANTSIZE2 20 /* 20 bit mantissa. */
417 #define EXPSIZE2 5 /* 5 bit base 2 exponent. */
418 #define MAXFRACT2 ((1ul << MANTSIZE2) - 1) /* Maximum fractional value. */
419 #define MAXEXP2 ((1 <<EXPSIZE2) - 1) /* Maximum exponent. */
420
encode_comp2_t(u64 value)421 static comp2_t encode_comp2_t(u64 value)
422 {
423 int exp, rnd;
424
425 exp = (value > (MAXFRACT2>>1));
426 rnd = 0;
427 while (value > MAXFRACT2) {
428 rnd = value & 1;
429 value >>= 1;
430 exp++;
431 }
432
433 /*
434 * If we need to round up, do it (and handle overflow correctly).
435 */
436 if (rnd && (++value > MAXFRACT2)) {
437 value >>= 1;
438 exp++;
439 }
440
441 if (exp > MAXEXP2) {
442 /* Overflow. Return largest representable number instead. */
443 return (1ul << (MANTSIZE2+EXPSIZE2-1)) - 1;
444 } else {
445 return (value & (MAXFRACT2>>1)) | (exp << (MANTSIZE2-1));
446 }
447 }
448 #endif
449
450 #if ACCT_VERSION==3
451 /*
452 * encode an u64 into a 32 bit IEEE float
453 */
encode_float(u64 value)454 static u32 encode_float(u64 value)
455 {
456 unsigned exp = 190;
457 unsigned u;
458
459 if (value==0) return 0;
460 while ((s64)value > 0){
461 value <<= 1;
462 exp--;
463 }
464 u = (u32)(value >> 40) & 0x7fffffu;
465 return u | (exp << 23);
466 }
467 #endif
468
469 /*
470 * Write an accounting entry for an exiting process
471 *
472 * The acct_process() call is the workhorse of the process
473 * accounting system. The struct acct is built here and then written
474 * into the accounting file. This function should only be called from
475 * do_exit() or when switching to a different output file.
476 */
477
478 /*
479 * do_acct_process does all actual work. Caller holds the reference to file.
480 */
do_acct_process(struct bsd_acct_struct * acct,struct pid_namespace * ns,struct file * file)481 static void do_acct_process(struct bsd_acct_struct *acct,
482 struct pid_namespace *ns, struct file *file)
483 {
484 struct pacct_struct *pacct = ¤t->signal->pacct;
485 acct_t ac;
486 mm_segment_t fs;
487 unsigned long flim;
488 u64 elapsed;
489 u64 run_time;
490 struct timespec uptime;
491 struct tty_struct *tty;
492
493 /*
494 * First check to see if there is enough free_space to continue
495 * the process accounting system.
496 */
497 if (!check_free_space(acct, file))
498 return;
499
500 /*
501 * Fill the accounting struct with the needed info as recorded
502 * by the different kernel functions.
503 */
504 memset((caddr_t)&ac, 0, sizeof(acct_t));
505
506 ac.ac_version = ACCT_VERSION | ACCT_BYTEORDER;
507 strlcpy(ac.ac_comm, current->comm, sizeof(ac.ac_comm));
508
509 /* calculate run_time in nsec*/
510 do_posix_clock_monotonic_gettime(&uptime);
511 run_time = (u64)uptime.tv_sec*NSEC_PER_SEC + uptime.tv_nsec;
512 run_time -= (u64)current->group_leader->start_time.tv_sec * NSEC_PER_SEC
513 + current->group_leader->start_time.tv_nsec;
514 /* convert nsec -> AHZ */
515 elapsed = nsec_to_AHZ(run_time);
516 #if ACCT_VERSION==3
517 ac.ac_etime = encode_float(elapsed);
518 #else
519 ac.ac_etime = encode_comp_t(elapsed < (unsigned long) -1l ?
520 (unsigned long) elapsed : (unsigned long) -1l);
521 #endif
522 #if ACCT_VERSION==1 || ACCT_VERSION==2
523 {
524 /* new enlarged etime field */
525 comp2_t etime = encode_comp2_t(elapsed);
526 ac.ac_etime_hi = etime >> 16;
527 ac.ac_etime_lo = (u16) etime;
528 }
529 #endif
530 do_div(elapsed, AHZ);
531 ac.ac_btime = get_seconds() - elapsed;
532 /* we really need to bite the bullet and change layout */
533 current_uid_gid(&ac.ac_uid, &ac.ac_gid);
534 #if ACCT_VERSION==2
535 ac.ac_ahz = AHZ;
536 #endif
537 #if ACCT_VERSION==1 || ACCT_VERSION==2
538 /* backward-compatible 16 bit fields */
539 ac.ac_uid16 = ac.ac_uid;
540 ac.ac_gid16 = ac.ac_gid;
541 #endif
542 #if ACCT_VERSION==3
543 ac.ac_pid = task_tgid_nr_ns(current, ns);
544 rcu_read_lock();
545 ac.ac_ppid = task_tgid_nr_ns(rcu_dereference(current->real_parent), ns);
546 rcu_read_unlock();
547 #endif
548
549 spin_lock_irq(¤t->sighand->siglock);
550 tty = current->signal->tty; /* Safe as we hold the siglock */
551 ac.ac_tty = tty ? old_encode_dev(tty_devnum(tty)) : 0;
552 ac.ac_utime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_utime)));
553 ac.ac_stime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_stime)));
554 ac.ac_flag = pacct->ac_flag;
555 ac.ac_mem = encode_comp_t(pacct->ac_mem);
556 ac.ac_minflt = encode_comp_t(pacct->ac_minflt);
557 ac.ac_majflt = encode_comp_t(pacct->ac_majflt);
558 ac.ac_exitcode = pacct->ac_exitcode;
559 spin_unlock_irq(¤t->sighand->siglock);
560 ac.ac_io = encode_comp_t(0 /* current->io_usage */); /* %% */
561 ac.ac_rw = encode_comp_t(ac.ac_io / 1024);
562 ac.ac_swaps = encode_comp_t(0);
563
564 /*
565 * Kernel segment override to datasegment and write it
566 * to the accounting file.
567 */
568 fs = get_fs();
569 set_fs(KERNEL_DS);
570 /*
571 * Accounting records are not subject to resource limits.
572 */
573 flim = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
574 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
575 file->f_op->write(file, (char *)&ac,
576 sizeof(acct_t), &file->f_pos);
577 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = flim;
578 set_fs(fs);
579 }
580
581 /**
582 * acct_init_pacct - initialize a new pacct_struct
583 * @pacct: per-process accounting info struct to initialize
584 */
acct_init_pacct(struct pacct_struct * pacct)585 void acct_init_pacct(struct pacct_struct *pacct)
586 {
587 memset(pacct, 0, sizeof(struct pacct_struct));
588 pacct->ac_utime = pacct->ac_stime = cputime_zero;
589 }
590
591 /**
592 * acct_collect - collect accounting information into pacct_struct
593 * @exitcode: task exit code
594 * @group_dead: not 0, if this thread is the last one in the process.
595 */
acct_collect(long exitcode,int group_dead)596 void acct_collect(long exitcode, int group_dead)
597 {
598 struct pacct_struct *pacct = ¤t->signal->pacct;
599 unsigned long vsize = 0;
600
601 if (group_dead && current->mm) {
602 struct vm_area_struct *vma;
603 down_read(¤t->mm->mmap_sem);
604 vma = current->mm->mmap;
605 while (vma) {
606 vsize += vma->vm_end - vma->vm_start;
607 vma = vma->vm_next;
608 }
609 up_read(¤t->mm->mmap_sem);
610 }
611
612 spin_lock_irq(¤t->sighand->siglock);
613 if (group_dead)
614 pacct->ac_mem = vsize / 1024;
615 if (thread_group_leader(current)) {
616 pacct->ac_exitcode = exitcode;
617 if (current->flags & PF_FORKNOEXEC)
618 pacct->ac_flag |= AFORK;
619 }
620 if (current->flags & PF_SUPERPRIV)
621 pacct->ac_flag |= ASU;
622 if (current->flags & PF_DUMPCORE)
623 pacct->ac_flag |= ACORE;
624 if (current->flags & PF_SIGNALED)
625 pacct->ac_flag |= AXSIG;
626 pacct->ac_utime = cputime_add(pacct->ac_utime, current->utime);
627 pacct->ac_stime = cputime_add(pacct->ac_stime, current->stime);
628 pacct->ac_minflt += current->min_flt;
629 pacct->ac_majflt += current->maj_flt;
630 spin_unlock_irq(¤t->sighand->siglock);
631 }
632
acct_process_in_ns(struct pid_namespace * ns)633 static void acct_process_in_ns(struct pid_namespace *ns)
634 {
635 struct file *file = NULL;
636 struct bsd_acct_struct *acct;
637
638 acct = ns->bacct;
639 /*
640 * accelerate the common fastpath:
641 */
642 if (!acct || !acct->file)
643 return;
644
645 spin_lock(&acct_lock);
646 file = acct->file;
647 if (unlikely(!file)) {
648 spin_unlock(&acct_lock);
649 return;
650 }
651 get_file(file);
652 spin_unlock(&acct_lock);
653
654 do_acct_process(acct, ns, file);
655 fput(file);
656 }
657
658 /**
659 * acct_process - now just a wrapper around acct_process_in_ns,
660 * which in turn is a wrapper around do_acct_process.
661 *
662 * handles process accounting for an exiting task
663 */
acct_process(void)664 void acct_process(void)
665 {
666 struct pid_namespace *ns;
667
668 /*
669 * This loop is safe lockless, since current is still
670 * alive and holds its namespace, which in turn holds
671 * its parent.
672 */
673 for (ns = task_active_pid_ns(current); ns != NULL; ns = ns->parent)
674 acct_process_in_ns(ns);
675 }
676