1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * linux/kernel/acct.c
4 *
5 * BSD Process Accounting for Linux
6 *
7 * Author: Marco van Wieringen <mvw@planets.elm.net>
8 *
9 * Some code based on ideas and code from:
10 * Thomas K. Dyas <tdyas@eden.rutgers.edu>
11 *
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.
17 *
18 * (C) Copyright 1995 - 1997 Marco van Wieringen - ELM Consultancy B.V.
19 *
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.
24 *
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.
27 *
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.
35 *
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.
41 *
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.
45 */
46
47 #include <linux/mm.h>
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>
61
62 #include <asm/div64.h>
63 #include <linux/pid_namespace.h>
64 #include <linux/fs_pin.h>
65
66 /*
67 * These constants control the amount of freespace that suspend and
68 * resume the process accounting system, and the time delay between
69 * each check.
70 * Turned into sysctl-controllable parameters. AV, 12/11/98
71 */
72
73 static int acct_parm[3] = {4, 2, 30};
74 #define RESUME (acct_parm[0]) /* >foo% free space - resume */
75 #define SUSPEND (acct_parm[1]) /* <foo% free space - suspend */
76 #define ACCT_TIMEOUT (acct_parm[2]) /* foo second timeout between checks */
77
78 #ifdef CONFIG_SYSCTL
79 static struct ctl_table kern_acct_table[] = {
80 {
81 .procname = "acct",
82 .data = &acct_parm,
83 .maxlen = 3*sizeof(int),
84 .mode = 0644,
85 .proc_handler = proc_dointvec,
86 },
87 { }
88 };
89
kernel_acct_sysctls_init(void)90 static __init int kernel_acct_sysctls_init(void)
91 {
92 register_sysctl_init("kernel", kern_acct_table);
93 return 0;
94 }
95 late_initcall(kernel_acct_sysctls_init);
96 #endif /* CONFIG_SYSCTL */
97
98 /*
99 * External references and all of the globals.
100 */
101
102 struct bsd_acct_struct {
103 struct fs_pin pin;
104 atomic_long_t count;
105 struct rcu_head rcu;
106 struct mutex lock;
107 int active;
108 unsigned long needcheck;
109 struct file *file;
110 struct pid_namespace *ns;
111 struct work_struct work;
112 struct completion done;
113 };
114
115 static void do_acct_process(struct bsd_acct_struct *acct);
116
117 /*
118 * Check the amount of free space and suspend/resume accordingly.
119 */
check_free_space(struct bsd_acct_struct * acct)120 static int check_free_space(struct bsd_acct_struct *acct)
121 {
122 struct kstatfs sbuf;
123
124 if (time_is_after_jiffies(acct->needcheck))
125 goto out;
126
127 /* May block */
128 if (vfs_statfs(&acct->file->f_path, &sbuf))
129 goto out;
130
131 if (acct->active) {
132 u64 suspend = sbuf.f_blocks * SUSPEND;
133 do_div(suspend, 100);
134 if (sbuf.f_bavail <= suspend) {
135 acct->active = 0;
136 pr_info("Process accounting paused\n");
137 }
138 } else {
139 u64 resume = sbuf.f_blocks * RESUME;
140 do_div(resume, 100);
141 if (sbuf.f_bavail >= resume) {
142 acct->active = 1;
143 pr_info("Process accounting resumed\n");
144 }
145 }
146
147 acct->needcheck = jiffies + ACCT_TIMEOUT*HZ;
148 out:
149 return acct->active;
150 }
151
acct_put(struct bsd_acct_struct * p)152 static void acct_put(struct bsd_acct_struct *p)
153 {
154 if (atomic_long_dec_and_test(&p->count))
155 kfree_rcu(p, rcu);
156 }
157
to_acct(struct fs_pin * p)158 static inline struct bsd_acct_struct *to_acct(struct fs_pin *p)
159 {
160 return p ? container_of(p, struct bsd_acct_struct, pin) : NULL;
161 }
162
acct_get(struct pid_namespace * ns)163 static struct bsd_acct_struct *acct_get(struct pid_namespace *ns)
164 {
165 struct bsd_acct_struct *res;
166 again:
167 smp_rmb();
168 rcu_read_lock();
169 res = to_acct(READ_ONCE(ns->bacct));
170 if (!res) {
171 rcu_read_unlock();
172 return NULL;
173 }
174 if (!atomic_long_inc_not_zero(&res->count)) {
175 rcu_read_unlock();
176 cpu_relax();
177 goto again;
178 }
179 rcu_read_unlock();
180 mutex_lock(&res->lock);
181 if (res != to_acct(READ_ONCE(ns->bacct))) {
182 mutex_unlock(&res->lock);
183 acct_put(res);
184 goto again;
185 }
186 return res;
187 }
188
acct_pin_kill(struct fs_pin * pin)189 static void acct_pin_kill(struct fs_pin *pin)
190 {
191 struct bsd_acct_struct *acct = to_acct(pin);
192 mutex_lock(&acct->lock);
193 do_acct_process(acct);
194 schedule_work(&acct->work);
195 wait_for_completion(&acct->done);
196 cmpxchg(&acct->ns->bacct, pin, NULL);
197 mutex_unlock(&acct->lock);
198 pin_remove(pin);
199 acct_put(acct);
200 }
201
close_work(struct work_struct * work)202 static void close_work(struct work_struct *work)
203 {
204 struct bsd_acct_struct *acct = container_of(work, struct bsd_acct_struct, work);
205 struct file *file = acct->file;
206 if (file->f_op->flush)
207 file->f_op->flush(file, NULL);
208 __fput_sync(file);
209 complete(&acct->done);
210 }
211
acct_on(struct filename * pathname)212 static int acct_on(struct filename *pathname)
213 {
214 struct file *file;
215 struct vfsmount *mnt, *internal;
216 struct pid_namespace *ns = task_active_pid_ns(current);
217 struct bsd_acct_struct *acct;
218 struct fs_pin *old;
219 int err;
220
221 acct = kzalloc(sizeof(struct bsd_acct_struct), GFP_KERNEL);
222 if (!acct)
223 return -ENOMEM;
224
225 /* Difference from BSD - they don't do O_APPEND */
226 file = file_open_name(pathname, O_WRONLY|O_APPEND|O_LARGEFILE, 0);
227 if (IS_ERR(file)) {
228 kfree(acct);
229 return PTR_ERR(file);
230 }
231
232 if (!S_ISREG(file_inode(file)->i_mode)) {
233 kfree(acct);
234 filp_close(file, NULL);
235 return -EACCES;
236 }
237
238 if (!(file->f_mode & FMODE_CAN_WRITE)) {
239 kfree(acct);
240 filp_close(file, NULL);
241 return -EIO;
242 }
243 internal = mnt_clone_internal(&file->f_path);
244 if (IS_ERR(internal)) {
245 kfree(acct);
246 filp_close(file, NULL);
247 return PTR_ERR(internal);
248 }
249 err = __mnt_want_write(internal);
250 if (err) {
251 mntput(internal);
252 kfree(acct);
253 filp_close(file, NULL);
254 return err;
255 }
256 mnt = file->f_path.mnt;
257 file->f_path.mnt = internal;
258
259 atomic_long_set(&acct->count, 1);
260 init_fs_pin(&acct->pin, acct_pin_kill);
261 acct->file = file;
262 acct->needcheck = jiffies;
263 acct->ns = ns;
264 mutex_init(&acct->lock);
265 INIT_WORK(&acct->work, close_work);
266 init_completion(&acct->done);
267 mutex_lock_nested(&acct->lock, 1); /* nobody has seen it yet */
268 pin_insert(&acct->pin, mnt);
269
270 rcu_read_lock();
271 old = xchg(&ns->bacct, &acct->pin);
272 mutex_unlock(&acct->lock);
273 pin_kill(old);
274 __mnt_drop_write(mnt);
275 mntput(mnt);
276 return 0;
277 }
278
279 static DEFINE_MUTEX(acct_on_mutex);
280
281 /**
282 * sys_acct - enable/disable process accounting
283 * @name: file name for accounting records or NULL to shutdown accounting
284 *
285 * sys_acct() is the only system call needed to implement process
286 * accounting. It takes the name of the file where accounting records
287 * should be written. If the filename is NULL, accounting will be
288 * shutdown.
289 *
290 * Returns: 0 for success or negative errno values for failure.
291 */
SYSCALL_DEFINE1(acct,const char __user *,name)292 SYSCALL_DEFINE1(acct, const char __user *, name)
293 {
294 int error = 0;
295
296 if (!capable(CAP_SYS_PACCT))
297 return -EPERM;
298
299 if (name) {
300 struct filename *tmp = getname(name);
301
302 if (IS_ERR(tmp))
303 return PTR_ERR(tmp);
304 mutex_lock(&acct_on_mutex);
305 error = acct_on(tmp);
306 mutex_unlock(&acct_on_mutex);
307 putname(tmp);
308 } else {
309 rcu_read_lock();
310 pin_kill(task_active_pid_ns(current)->bacct);
311 }
312
313 return error;
314 }
315
acct_exit_ns(struct pid_namespace * ns)316 void acct_exit_ns(struct pid_namespace *ns)
317 {
318 rcu_read_lock();
319 pin_kill(ns->bacct);
320 }
321
322 /*
323 * encode an u64 into a comp_t
324 *
325 * This routine has been adopted from the encode_comp_t() function in
326 * the kern_acct.c file of the FreeBSD operating system. The encoding
327 * is a 13-bit fraction with a 3-bit (base 8) exponent.
328 */
329
330 #define MANTSIZE 13 /* 13 bit mantissa. */
331 #define EXPSIZE 3 /* Base 8 (3 bit) exponent. */
332 #define MAXFRACT ((1 << MANTSIZE) - 1) /* Maximum fractional value. */
333
encode_comp_t(u64 value)334 static comp_t encode_comp_t(u64 value)
335 {
336 int exp, rnd;
337
338 exp = rnd = 0;
339 while (value > MAXFRACT) {
340 rnd = value & (1 << (EXPSIZE - 1)); /* Round up? */
341 value >>= EXPSIZE; /* Base 8 exponent == 3 bit shift. */
342 exp++;
343 }
344
345 /*
346 * If we need to round up, do it (and handle overflow correctly).
347 */
348 if (rnd && (++value > MAXFRACT)) {
349 value >>= EXPSIZE;
350 exp++;
351 }
352
353 if (exp > (((comp_t) ~0U) >> MANTSIZE))
354 return (comp_t) ~0U;
355 /*
356 * Clean it up and polish it off.
357 */
358 exp <<= MANTSIZE; /* Shift the exponent into place */
359 exp += value; /* and add on the mantissa. */
360 return exp;
361 }
362
363 #if ACCT_VERSION == 1 || ACCT_VERSION == 2
364 /*
365 * encode an u64 into a comp2_t (24 bits)
366 *
367 * Format: 5 bit base 2 exponent, 20 bits mantissa.
368 * The leading bit of the mantissa is not stored, but implied for
369 * non-zero exponents.
370 * Largest encodable value is 50 bits.
371 */
372
373 #define MANTSIZE2 20 /* 20 bit mantissa. */
374 #define EXPSIZE2 5 /* 5 bit base 2 exponent. */
375 #define MAXFRACT2 ((1ul << MANTSIZE2) - 1) /* Maximum fractional value. */
376 #define MAXEXP2 ((1 << EXPSIZE2) - 1) /* Maximum exponent. */
377
encode_comp2_t(u64 value)378 static comp2_t encode_comp2_t(u64 value)
379 {
380 int exp, rnd;
381
382 exp = (value > (MAXFRACT2>>1));
383 rnd = 0;
384 while (value > MAXFRACT2) {
385 rnd = value & 1;
386 value >>= 1;
387 exp++;
388 }
389
390 /*
391 * If we need to round up, do it (and handle overflow correctly).
392 */
393 if (rnd && (++value > MAXFRACT2)) {
394 value >>= 1;
395 exp++;
396 }
397
398 if (exp > MAXEXP2) {
399 /* Overflow. Return largest representable number instead. */
400 return (1ul << (MANTSIZE2+EXPSIZE2-1)) - 1;
401 } else {
402 return (value & (MAXFRACT2>>1)) | (exp << (MANTSIZE2-1));
403 }
404 }
405 #elif ACCT_VERSION == 3
406 /*
407 * encode an u64 into a 32 bit IEEE float
408 */
encode_float(u64 value)409 static u32 encode_float(u64 value)
410 {
411 unsigned exp = 190;
412 unsigned u;
413
414 if (value == 0)
415 return 0;
416 while ((s64)value > 0) {
417 value <<= 1;
418 exp--;
419 }
420 u = (u32)(value >> 40) & 0x7fffffu;
421 return u | (exp << 23);
422 }
423 #endif
424
425 /*
426 * Write an accounting entry for an exiting process
427 *
428 * The acct_process() call is the workhorse of the process
429 * accounting system. The struct acct is built here and then written
430 * into the accounting file. This function should only be called from
431 * do_exit() or when switching to a different output file.
432 */
433
fill_ac(acct_t * ac)434 static void fill_ac(acct_t *ac)
435 {
436 struct pacct_struct *pacct = ¤t->signal->pacct;
437 u64 elapsed, run_time;
438 time64_t btime;
439 struct tty_struct *tty;
440
441 /*
442 * Fill the accounting struct with the needed info as recorded
443 * by the different kernel functions.
444 */
445 memset(ac, 0, sizeof(acct_t));
446
447 ac->ac_version = ACCT_VERSION | ACCT_BYTEORDER;
448 strscpy(ac->ac_comm, current->comm, sizeof(ac->ac_comm));
449
450 /* calculate run_time in nsec*/
451 run_time = ktime_get_ns();
452 run_time -= current->group_leader->start_time;
453 /* convert nsec -> AHZ */
454 elapsed = nsec_to_AHZ(run_time);
455 #if ACCT_VERSION == 3
456 ac->ac_etime = encode_float(elapsed);
457 #else
458 ac->ac_etime = encode_comp_t(elapsed < (unsigned long) -1l ?
459 (unsigned long) elapsed : (unsigned long) -1l);
460 #endif
461 #if ACCT_VERSION == 1 || ACCT_VERSION == 2
462 {
463 /* new enlarged etime field */
464 comp2_t etime = encode_comp2_t(elapsed);
465
466 ac->ac_etime_hi = etime >> 16;
467 ac->ac_etime_lo = (u16) etime;
468 }
469 #endif
470 do_div(elapsed, AHZ);
471 btime = ktime_get_real_seconds() - elapsed;
472 ac->ac_btime = clamp_t(time64_t, btime, 0, U32_MAX);
473 #if ACCT_VERSION == 2
474 ac->ac_ahz = AHZ;
475 #endif
476
477 spin_lock_irq(¤t->sighand->siglock);
478 tty = current->signal->tty; /* Safe as we hold the siglock */
479 ac->ac_tty = tty ? old_encode_dev(tty_devnum(tty)) : 0;
480 ac->ac_utime = encode_comp_t(nsec_to_AHZ(pacct->ac_utime));
481 ac->ac_stime = encode_comp_t(nsec_to_AHZ(pacct->ac_stime));
482 ac->ac_flag = pacct->ac_flag;
483 ac->ac_mem = encode_comp_t(pacct->ac_mem);
484 ac->ac_minflt = encode_comp_t(pacct->ac_minflt);
485 ac->ac_majflt = encode_comp_t(pacct->ac_majflt);
486 ac->ac_exitcode = pacct->ac_exitcode;
487 spin_unlock_irq(¤t->sighand->siglock);
488 }
489 /*
490 * do_acct_process does all actual work. Caller holds the reference to file.
491 */
do_acct_process(struct bsd_acct_struct * acct)492 static void do_acct_process(struct bsd_acct_struct *acct)
493 {
494 acct_t ac;
495 unsigned long flim;
496 const struct cred *orig_cred;
497 struct file *file = acct->file;
498
499 /*
500 * Accounting records are not subject to resource limits.
501 */
502 flim = rlimit(RLIMIT_FSIZE);
503 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
504 /* Perform file operations on behalf of whoever enabled accounting */
505 orig_cred = override_creds(file->f_cred);
506
507 /*
508 * First check to see if there is enough free_space to continue
509 * the process accounting system.
510 */
511 if (!check_free_space(acct))
512 goto out;
513
514 fill_ac(&ac);
515 /* we really need to bite the bullet and change layout */
516 ac.ac_uid = from_kuid_munged(file->f_cred->user_ns, orig_cred->uid);
517 ac.ac_gid = from_kgid_munged(file->f_cred->user_ns, orig_cred->gid);
518 #if ACCT_VERSION == 1 || ACCT_VERSION == 2
519 /* backward-compatible 16 bit fields */
520 ac.ac_uid16 = ac.ac_uid;
521 ac.ac_gid16 = ac.ac_gid;
522 #elif ACCT_VERSION == 3
523 {
524 struct pid_namespace *ns = acct->ns;
525
526 ac.ac_pid = task_tgid_nr_ns(current, ns);
527 rcu_read_lock();
528 ac.ac_ppid = task_tgid_nr_ns(rcu_dereference(current->real_parent),
529 ns);
530 rcu_read_unlock();
531 }
532 #endif
533 /*
534 * Get freeze protection. If the fs is frozen, just skip the write
535 * as we could deadlock the system otherwise.
536 */
537 if (file_start_write_trylock(file)) {
538 /* it's been opened O_APPEND, so position is irrelevant */
539 loff_t pos = 0;
540 __kernel_write(file, &ac, sizeof(acct_t), &pos);
541 file_end_write(file);
542 }
543 out:
544 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = flim;
545 revert_creds(orig_cred);
546 }
547
548 /**
549 * acct_collect - collect accounting information into pacct_struct
550 * @exitcode: task exit code
551 * @group_dead: not 0, if this thread is the last one in the process.
552 */
acct_collect(long exitcode,int group_dead)553 void acct_collect(long exitcode, int group_dead)
554 {
555 struct pacct_struct *pacct = ¤t->signal->pacct;
556 u64 utime, stime;
557 unsigned long vsize = 0;
558
559 if (group_dead && current->mm) {
560 struct mm_struct *mm = current->mm;
561 VMA_ITERATOR(vmi, mm, 0);
562 struct vm_area_struct *vma;
563
564 mmap_read_lock(mm);
565 for_each_vma(vmi, vma)
566 vsize += vma->vm_end - vma->vm_start;
567 mmap_read_unlock(mm);
568 }
569
570 spin_lock_irq(¤t->sighand->siglock);
571 if (group_dead)
572 pacct->ac_mem = vsize / 1024;
573 if (thread_group_leader(current)) {
574 pacct->ac_exitcode = exitcode;
575 if (current->flags & PF_FORKNOEXEC)
576 pacct->ac_flag |= AFORK;
577 }
578 if (current->flags & PF_SUPERPRIV)
579 pacct->ac_flag |= ASU;
580 if (current->flags & PF_DUMPCORE)
581 pacct->ac_flag |= ACORE;
582 if (current->flags & PF_SIGNALED)
583 pacct->ac_flag |= AXSIG;
584
585 task_cputime(current, &utime, &stime);
586 pacct->ac_utime += utime;
587 pacct->ac_stime += stime;
588 pacct->ac_minflt += current->min_flt;
589 pacct->ac_majflt += current->maj_flt;
590 spin_unlock_irq(¤t->sighand->siglock);
591 }
592
slow_acct_process(struct pid_namespace * ns)593 static void slow_acct_process(struct pid_namespace *ns)
594 {
595 for ( ; ns; ns = ns->parent) {
596 struct bsd_acct_struct *acct = acct_get(ns);
597 if (acct) {
598 do_acct_process(acct);
599 mutex_unlock(&acct->lock);
600 acct_put(acct);
601 }
602 }
603 }
604
605 /**
606 * acct_process - handles process accounting for an exiting task
607 */
acct_process(void)608 void acct_process(void)
609 {
610 struct pid_namespace *ns;
611
612 /*
613 * This loop is safe lockless, since current is still
614 * alive and holds its namespace, which in turn holds
615 * its parent.
616 */
617 for (ns = task_active_pid_ns(current); ns != NULL; ns = ns->parent) {
618 if (ns->bacct)
619 break;
620 }
621 if (unlikely(ns))
622 slow_acct_process(ns);
623 }
624