projects/linux·blob v2.6.34 · refs

675 lines · 17.2 KiB

1/*
2 * linux/kernel/acct.c
3 *
4 * BSD Process Accounting for Linux
5 *
6 * Author: Marco van Wieringen <[email protected]>
7 *
8 * Some code based on ideas and code from:
9 * Thomas K. Dyas <[email protected]>
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
70int 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 */
78static 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 */
86struct 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
95static DEFINE_SPINLOCK(acct_lock);
96static LIST_HEAD(acct_list);
97
98/*
99 * Called whenever the timer says to check the free space.
100 */
101static 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 */
110static 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;
168out:
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 */
179static 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
215static int acct_on(char *name)
216{
217 struct file *file;
218 struct vfsmount *mnt;
219 int error;
220 struct pid_namespace *ns;
221 struct bsd_acct_struct *acct = NULL;
222
223 /* Difference from BSD - they don't do O_APPEND */
224 file = filp_open(name, O_WRONLY|O_APPEND|O_LARGEFILE, 0);
225 if (IS_ERR(file))
226 return PTR_ERR(file);
227
228 if (!S_ISREG(file->f_path.dentry->d_inode->i_mode)) {
229 filp_close(file, NULL);
230 return -EACCES;
231 }
232
233 if (!file->f_op->write) {
234 filp_close(file, NULL);
235 return -EIO;
236 }
237
238 ns = task_active_pid_ns(current);
239 if (ns->bacct == NULL) {
240 acct = kzalloc(sizeof(struct bsd_acct_struct), GFP_KERNEL);
241 if (acct == NULL) {
242 filp_close(file, NULL);
243 return -ENOMEM;
244 }
245 }
246
247 error = security_acct(file);
248 if (error) {
249 kfree(acct);
250 filp_close(file, NULL);
251 return error;
252 }
253
254 spin_lock(&acct_lock);
255 if (ns->bacct == NULL) {
256 ns->bacct = acct;
257 acct = NULL;
258 }
259
260 mnt = file->f_path.mnt;
261 mnt_pin(mnt);
262 acct_file_reopen(ns->bacct, file, ns);
263 spin_unlock(&acct_lock);
264
265 mntput(mnt); /* it's pinned, now give up active reference */
266 kfree(acct);
267
268 return 0;
269}
270
271/**
272 * sys_acct - enable/disable process accounting
273 * @name: file name for accounting records or NULL to shutdown accounting
274 *
275 * Returns 0 for success or negative errno values for failure.
276 *
277 * sys_acct() is the only system call needed to implement process
278 * accounting. It takes the name of the file where accounting records
279 * should be written. If the filename is NULL, accounting will be
280 * shutdown.
281 */
282SYSCALL_DEFINE1(acct, const char __user *, name)
283{
284 int error;
285
286 if (!capable(CAP_SYS_PACCT))
287 return -EPERM;
288
289 if (name) {
290 char *tmp = getname(name);
291 if (IS_ERR(tmp))
292 return (PTR_ERR(tmp));
293 error = acct_on(tmp);
294 putname(tmp);
295 } else {
296 struct bsd_acct_struct *acct;
297
298 acct = task_active_pid_ns(current)->bacct;
299 if (acct == NULL)
300 return 0;
301
302 error = security_acct(NULL);
303 if (!error) {
304 spin_lock(&acct_lock);
305 acct_file_reopen(acct, NULL, NULL);
306 spin_unlock(&acct_lock);
307 }
308 }
309 return error;
310}
311
312/**
313 * acct_auto_close - turn off a filesystem's accounting if it is on
314 * @m: vfsmount being shut down
315 *
316 * If the accounting is turned on for a file in the subtree pointed to
317 * to by m, turn accounting off. Done when m is about to die.
318 */
319void acct_auto_close_mnt(struct vfsmount *m)
320{
321 struct bsd_acct_struct *acct;
322
323 spin_lock(&acct_lock);
324restart:
325 list_for_each_entry(acct, &acct_list, list)
326 if (acct->file && acct->file->f_path.mnt == m) {
327 acct_file_reopen(acct, NULL, NULL);
328 goto restart;
329 }
330 spin_unlock(&acct_lock);
331}
332
333/**
334 * acct_auto_close - turn off a filesystem's accounting if it is on
335 * @sb: super block for the filesystem
336 *
337 * If the accounting is turned on for a file in the filesystem pointed
338 * to by sb, turn accounting off.
339 */
340void acct_auto_close(struct super_block *sb)
341{
342 struct bsd_acct_struct *acct;
343
344 spin_lock(&acct_lock);
345restart:
346 list_for_each_entry(acct, &acct_list, list)
347 if (acct->file && acct->file->f_path.mnt->mnt_sb == sb) {
348 acct_file_reopen(acct, NULL, NULL);
349 goto restart;
350 }
351 spin_unlock(&acct_lock);
352}
353
354void acct_exit_ns(struct pid_namespace *ns)
355{
356 struct bsd_acct_struct *acct = ns->bacct;
357
358 if (acct == NULL)
359 return;
360
361 del_timer_sync(&acct->timer);
362 spin_lock(&acct_lock);
363 if (acct->file != NULL)
364 acct_file_reopen(acct, NULL, NULL);
365 spin_unlock(&acct_lock);
366
367 kfree(acct);
368}
369
370/*
371 * encode an unsigned long into a comp_t
372 *
373 * This routine has been adopted from the encode_comp_t() function in
374 * the kern_acct.c file of the FreeBSD operating system. The encoding
375 * is a 13-bit fraction with a 3-bit (base 8) exponent.
376 */
377
378#define MANTSIZE 13 /* 13 bit mantissa. */
379#define EXPSIZE 3 /* Base 8 (3 bit) exponent. */
380#define MAXFRACT ((1 << MANTSIZE) - 1) /* Maximum fractional value. */
381
382static comp_t encode_comp_t(unsigned long value)
383{
384 int exp, rnd;
385
386 exp = rnd = 0;
387 while (value > MAXFRACT) {
388 rnd = value & (1 << (EXPSIZE - 1)); /* Round up? */
389 value >>= EXPSIZE; /* Base 8 exponent == 3 bit shift. */
390 exp++;
391 }
392
393 /*
394 * If we need to round up, do it (and handle overflow correctly).
395 */
396 if (rnd && (++value > MAXFRACT)) {
397 value >>= EXPSIZE;
398 exp++;
399 }
400
401 /*
402 * Clean it up and polish it off.
403 */
404 exp <<= MANTSIZE; /* Shift the exponent into place */
405 exp += value; /* and add on the mantissa. */
406 return exp;
407}
408
409#if ACCT_VERSION==1 || ACCT_VERSION==2
410/*
411 * encode an u64 into a comp2_t (24 bits)
412 *
413 * Format: 5 bit base 2 exponent, 20 bits mantissa.
414 * The leading bit of the mantissa is not stored, but implied for
415 * non-zero exponents.
416 * Largest encodable value is 50 bits.
417 */
418
419#define MANTSIZE2 20 /* 20 bit mantissa. */
420#define EXPSIZE2 5 /* 5 bit base 2 exponent. */
421#define MAXFRACT2 ((1ul << MANTSIZE2) - 1) /* Maximum fractional value. */
422#define MAXEXP2 ((1 <<EXPSIZE2) - 1) /* Maximum exponent. */
423
424static comp2_t encode_comp2_t(u64 value)
425{
426 int exp, rnd;
427
428 exp = (value > (MAXFRACT2>>1));
429 rnd = 0;
430 while (value > MAXFRACT2) {
431 rnd = value & 1;
432 value >>= 1;
433 exp++;
434 }
435
436 /*
437 * If we need to round up, do it (and handle overflow correctly).
438 */
439 if (rnd && (++value > MAXFRACT2)) {
440 value >>= 1;
441 exp++;
442 }
443
444 if (exp > MAXEXP2) {
445 /* Overflow. Return largest representable number instead. */
446 return (1ul << (MANTSIZE2+EXPSIZE2-1)) - 1;
447 } else {
448 return (value & (MAXFRACT2>>1)) | (exp << (MANTSIZE2-1));
449 }
450}
451#endif
452
453#if ACCT_VERSION==3
454/*
455 * encode an u64 into a 32 bit IEEE float
456 */
457static u32 encode_float(u64 value)
458{
459 unsigned exp = 190;
460 unsigned u;
461
462 if (value==0) return 0;
463 while ((s64)value > 0){
464 value <<= 1;
465 exp--;
466 }
467 u = (u32)(value >> 40) & 0x7fffffu;
468 return u | (exp << 23);
469}
470#endif
471
472/*
473 * Write an accounting entry for an exiting process
474 *
475 * The acct_process() call is the workhorse of the process
476 * accounting system. The struct acct is built here and then written
477 * into the accounting file. This function should only be called from
478 * do_exit() or when switching to a different output file.
479 */
480
481/*
482 * do_acct_process does all actual work. Caller holds the reference to file.
483 */
484static void do_acct_process(struct bsd_acct_struct *acct,
485 struct pid_namespace *ns, struct file *file)
486{
487 struct pacct_struct *pacct = &current->signal->pacct;
488 acct_t ac;
489 mm_segment_t fs;
490 unsigned long flim;
491 u64 elapsed;
492 u64 run_time;
493 struct timespec uptime;
494 struct tty_struct *tty;
495 const struct cred *orig_cred;
496
497 /* Perform file operations on behalf of whoever enabled accounting */
498 orig_cred = override_creds(file->f_cred);
499
500 /*
501 * First check to see if there is enough free_space to continue
502 * the process accounting system.
503 */
504 if (!check_free_space(acct, file))
505 goto out;
506
507 /*
508 * Fill the accounting struct with the needed info as recorded
509 * by the different kernel functions.
510 */
511 memset((caddr_t)&ac, 0, sizeof(acct_t));
512
513 ac.ac_version = ACCT_VERSION | ACCT_BYTEORDER;
514 strlcpy(ac.ac_comm, current->comm, sizeof(ac.ac_comm));
515
516 /* calculate run_time in nsec*/
517 do_posix_clock_monotonic_gettime(&uptime);
518 run_time = (u64)uptime.tv_sec*NSEC_PER_SEC + uptime.tv_nsec;
519 run_time -= (u64)current->group_leader->start_time.tv_sec * NSEC_PER_SEC
520 + current->group_leader->start_time.tv_nsec;
521 /* convert nsec -> AHZ */
522 elapsed = nsec_to_AHZ(run_time);
523#if ACCT_VERSION==3
524 ac.ac_etime = encode_float(elapsed);
525#else
526 ac.ac_etime = encode_comp_t(elapsed < (unsigned long) -1l ?
527 (unsigned long) elapsed : (unsigned long) -1l);
528#endif
529#if ACCT_VERSION==1 || ACCT_VERSION==2
530 {
531 /* new enlarged etime field */
532 comp2_t etime = encode_comp2_t(elapsed);
533 ac.ac_etime_hi = etime >> 16;
534 ac.ac_etime_lo = (u16) etime;
535 }
536#endif
537 do_div(elapsed, AHZ);
538 ac.ac_btime = get_seconds() - elapsed;
539 /* we really need to bite the bullet and change layout */
540 ac.ac_uid = orig_cred->uid;
541 ac.ac_gid = orig_cred->gid;
542#if ACCT_VERSION==2
543 ac.ac_ahz = AHZ;
544#endif
545#if ACCT_VERSION==1 || ACCT_VERSION==2
546 /* backward-compatible 16 bit fields */
547 ac.ac_uid16 = ac.ac_uid;
548 ac.ac_gid16 = ac.ac_gid;
549#endif
550#if ACCT_VERSION==3
551 ac.ac_pid = task_tgid_nr_ns(current, ns);
552 rcu_read_lock();
553 ac.ac_ppid = task_tgid_nr_ns(rcu_dereference(current->real_parent), ns);
554 rcu_read_unlock();
555#endif
556
557 spin_lock_irq(&current->sighand->siglock);
558 tty = current->signal->tty; /* Safe as we hold the siglock */
559 ac.ac_tty = tty ? old_encode_dev(tty_devnum(tty)) : 0;
560 ac.ac_utime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_utime)));
561 ac.ac_stime = encode_comp_t(jiffies_to_AHZ(cputime_to_jiffies(pacct->ac_stime)));
562 ac.ac_flag = pacct->ac_flag;
563 ac.ac_mem = encode_comp_t(pacct->ac_mem);
564 ac.ac_minflt = encode_comp_t(pacct->ac_minflt);
565 ac.ac_majflt = encode_comp_t(pacct->ac_majflt);
566 ac.ac_exitcode = pacct->ac_exitcode;
567 spin_unlock_irq(&current->sighand->siglock);
568 ac.ac_io = encode_comp_t(0 /* current->io_usage */); /* %% */
569 ac.ac_rw = encode_comp_t(ac.ac_io / 1024);
570 ac.ac_swaps = encode_comp_t(0);
571
572 /*
573 * Kernel segment override to datasegment and write it
574 * to the accounting file.
575 */
576 fs = get_fs();
577 set_fs(KERNEL_DS);
578 /*
579 * Accounting records are not subject to resource limits.
580 */
581 flim = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
582 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
583 file->f_op->write(file, (char *)&ac,
584 sizeof(acct_t), &file->f_pos);
585 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = flim;
586 set_fs(fs);
587out:
588 revert_creds(orig_cred);
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 */
596void acct_collect(long exitcode, int group_dead)
597{
598 struct pacct_struct *pacct = &current->signal->pacct;
599 unsigned long vsize = 0;
600
601 if (group_dead && current->mm) {
602 struct vm_area_struct *vma;
603 down_read(&current->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(&current->mm->mmap_sem);
610 }
611
612 spin_lock_irq(&current->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(&current->sighand->siglock);
631}
632
633static 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 */
664void 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}