projects/linux·blob v6.13 · refs

622 lines · 15.9 KiB

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