]> git.proxmox.com Git - mirror_lxcfs.git/blob - bindings.c
lxcfs: add --pidfd option
[mirror_lxcfs.git] / bindings.c
1 /* lxcfs
2 *
3 * Copyright © 2014-2016 Canonical, Inc
4 * Author: Serge Hallyn <serge.hallyn@ubuntu.com>
5 *
6 * See COPYING file for details.
7 */
8
9 #ifndef _GNU_SOURCE
10 #define _GNU_SOURCE
11 #endif
12
13 #ifndef FUSE_USE_VERSION
14 #define FUSE_USE_VERSION 26
15 #endif
16
17 #define _FILE_OFFSET_BITS 64
18
19 #include <dirent.h>
20 #include <errno.h>
21 #include <fcntl.h>
22 #include <fuse.h>
23 #include <inttypes.h>
24 #include <libgen.h>
25 #include <pthread.h>
26 #include <sched.h>
27 #include <stdarg.h>
28 #include <stdbool.h>
29 #include <stdint.h>
30 #include <stdio.h>
31 #include <stdlib.h>
32 #include <string.h>
33 #include <time.h>
34 #include <unistd.h>
35 #include <wait.h>
36 #include <linux/magic.h>
37 #include <linux/sched.h>
38 #include <sys/epoll.h>
39 #include <sys/mman.h>
40 #include <sys/mount.h>
41 #include <sys/param.h>
42 #include <sys/socket.h>
43 #include <sys/syscall.h>
44 #include <sys/sysinfo.h>
45 #include <sys/vfs.h>
46
47 #include "bindings.h"
48 #include "config.h"
49 #include "cgroup_fuse.h"
50 #include "cgroups/cgroup.h"
51 #include "cgroups/cgroup_utils.h"
52 #include "memory_utils.h"
53 #include "proc_cpuview.h"
54 #include "utils.h"
55
56 static bool can_use_pidfd;
57
58 /* Define pivot_root() if missing from the C library */
59 #ifndef HAVE_PIVOT_ROOT
60 static int pivot_root(const char *new_root, const char *put_old)
61 {
62 #ifdef __NR_pivot_root
63 return syscall(__NR_pivot_root, new_root, put_old);
64 #else
65 errno = ENOSYS;
66 return -1;
67 #endif
68 }
69 #else
70 extern int pivot_root(const char *new_root, const char *put_old);
71 #endif
72
73 /*
74 * A table caching which pid is init for a pid namespace.
75 * When looking up which pid is init for $qpid, we first
76 * 1. Stat /proc/$qpid/ns/pid.
77 * 2. Check whether the ino_t is in our store.
78 * a. if not, fork a child in qpid's ns to send us
79 * ucred.pid = 1, and read the initpid. Cache
80 * initpid and creation time for /proc/initpid
81 * in a new store entry.
82 * b. if so, verify that /proc/initpid still matches
83 * what we have saved. If not, clear the store
84 * entry and go back to a. If so, return the
85 * cached initpid.
86 */
87 struct pidns_init_store {
88 ino_t ino; /* inode number for /proc/$pid/ns/pid */
89 pid_t initpid; /* the pid of nit in that ns */
90 int init_pidfd;
91 long int ctime; /* the time at which /proc/$initpid was created */
92 struct pidns_init_store *next;
93 long int lastcheck;
94 };
95
96 /* lol - look at how they are allocated in the kernel */
97 #define PIDNS_HASH_SIZE 4096
98 #define HASH(x) ((x) % PIDNS_HASH_SIZE)
99
100 static struct pidns_init_store *pidns_hash_table[PIDNS_HASH_SIZE];
101 static pthread_mutex_t pidns_store_mutex = PTHREAD_MUTEX_INITIALIZER;
102
103 static void lock_mutex(pthread_mutex_t *l)
104 {
105 int ret;
106
107 ret = pthread_mutex_lock(l);
108 if (ret)
109 log_exit("%s - returned %d\n", strerror(ret), ret);
110 }
111
112 struct cgroup_ops *cgroup_ops;
113
114 static void unlock_mutex(pthread_mutex_t *l)
115 {
116 int ret;
117
118 ret = pthread_mutex_unlock(l);
119 if (ret)
120 log_exit("%s - returned %d\n", strerror(ret), ret);
121 }
122
123 static void store_lock(void)
124 {
125 lock_mutex(&pidns_store_mutex);
126 }
127
128 static void store_unlock(void)
129 {
130 unlock_mutex(&pidns_store_mutex);
131 }
132
133 /* /proc/ = 6
134 * +
135 * <pid-as-str> = INTTYPE_TO_STRLEN(pid_t)
136 * +
137 * \0 = 1
138 */
139 #define LXCFS_PROC_PID_LEN \
140 (STRLITERALLEN("/proc/") + INTTYPE_TO_STRLEN(uint64_t) + +1)
141
142 /* Must be called under store_lock */
143 static bool initpid_still_valid(struct pidns_init_store *entry)
144 {
145 bool valid = true;
146
147 if (entry->init_pidfd >= 0) {
148 if (pidfd_send_signal(entry->init_pidfd, 0, NULL, 0))
149 valid = false;
150 } else {
151 struct stat st;
152 char path[LXCFS_PROC_PID_LEN];
153
154 snprintf(path, sizeof(path), "/proc/%d", entry->initpid);
155
156 if (stat(path, &st) || entry->ctime != st.st_ctime)
157 valid = false;
158 }
159
160 return valid;
161 }
162
163 /* Must be called under store_lock */
164 static void remove_initpid(struct pidns_init_store *entry)
165 {
166 struct pidns_init_store *it;
167 int ino_hash;
168
169 lxcfs_debug("Removing cached entry for pid %d from init pid cache",
170 entry->initpid);
171
172 ino_hash = HASH(entry->ino);
173 if (pidns_hash_table[ino_hash] == entry) {
174 pidns_hash_table[ino_hash] = entry->next;
175 close_prot_errno_disarm(entry->init_pidfd);
176 free_disarm(entry);
177 return;
178 }
179
180 it = pidns_hash_table[ino_hash];
181 while (it) {
182 if (it->next == entry) {
183 it->next = entry->next;
184 close_prot_errno_disarm(entry->init_pidfd);
185 free_disarm(entry);
186 return;
187 }
188 it = it->next;
189 }
190 }
191
192 #define PURGE_SECS 5
193 /* Must be called under store_lock */
194 static void prune_initpid_store(void)
195 {
196 static long int last_prune = 0;
197 long int now, threshold;
198
199 if (!last_prune) {
200 last_prune = time(NULL);
201 return;
202 }
203
204 now = time(NULL);
205 if (now < last_prune + PURGE_SECS)
206 return;
207
208 lxcfs_debug("Pruning init pid cache");
209
210 last_prune = now;
211 threshold = now - 2 * PURGE_SECS;
212
213 for (int i = 0; i < PIDNS_HASH_SIZE; i++) {
214 for (struct pidns_init_store *entry = pidns_hash_table[i], *prev = NULL; entry;) {
215 if (entry->lastcheck < threshold) {
216 struct pidns_init_store *cur = entry;
217
218 lxcfs_debug("Removed cache entry for pid %d to init pid cache", cur->initpid);
219
220 if (prev)
221 prev->next = entry->next;
222 else
223 pidns_hash_table[i] = entry->next;
224 entry = entry->next;
225 close_prot_errno_disarm(cur->init_pidfd);
226 free_disarm(cur);
227 } else {
228 prev = entry;
229 entry = entry->next;
230 }
231 }
232 }
233 }
234
235 /* Must be called under store_lock */
236 static void save_initpid(struct stat *sb, pid_t pid)
237 {
238 __do_free struct pidns_init_store *e = NULL;
239 __do_close_prot_errno int pidfd = -EBADF;
240 char path[LXCFS_PROC_PID_LEN];
241 struct lxcfs_opts *opts = fuse_get_context()->private_data;
242 struct stat st;
243 int ino_hash;
244
245 if (opts->use_pidfd && can_use_pidfd) {
246 pidfd = pidfd_open(pid, 0);
247 if (pidfd < 0)
248 return;
249 }
250
251 snprintf(path, sizeof(path), "/proc/%d", pid);
252 if (stat(path, &st))
253 return;
254
255 e = malloc(sizeof(*e));
256 if (!e)
257 return;
258
259 e->ino = sb->st_ino;
260 e->initpid = pid;
261 e->ctime = st.st_ctime;
262 ino_hash = HASH(e->ino);
263 e->next = pidns_hash_table[ino_hash];
264 e->lastcheck = time(NULL);
265 e->init_pidfd = move_fd(pidfd);
266 pidns_hash_table[ino_hash] = move_ptr(e);
267
268 lxcfs_debug("Added cache entry %d for pid %d to init pid cache", ino_hash, pid);
269 }
270
271 /*
272 * Given the stat(2) info for a nsfd pid inode, lookup the init_pid_store
273 * entry for the inode number and creation time. Verify that the init pid
274 * is still valid. If not, remove it. Return the entry if valid, NULL
275 * otherwise.
276 * Must be called under store_lock
277 */
278 static struct pidns_init_store *lookup_verify_initpid(struct stat *sb)
279 {
280 struct pidns_init_store *entry = pidns_hash_table[HASH(sb->st_ino)];
281
282 while (entry) {
283 if (entry->ino == sb->st_ino) {
284 if (initpid_still_valid(entry)) {
285 entry->lastcheck = time(NULL);
286 return entry;
287 }
288
289 remove_initpid(entry);
290 return NULL;
291 }
292 entry = entry->next;
293 }
294
295 return NULL;
296 }
297
298 static int send_creds_clone_wrapper(void *arg)
299 {
300 struct ucred cred;
301 char v;
302 int sock = *(int *)arg;
303
304 /* we are the child */
305 cred.uid = 0;
306 cred.gid = 0;
307 cred.pid = 1;
308 v = '1';
309 if (send_creds(sock, &cred, v, true) != SEND_CREDS_OK)
310 return 1;
311 return 0;
312 }
313
314 /*
315 * clone a task which switches to @task's namespace and writes '1'.
316 * over a unix sock so we can read the task's reaper's pid in our
317 * namespace
318 *
319 * Note: glibc's fork() does not respect pidns, which can lead to failed
320 * assertions inside glibc (and thus failed forks) if the child's pid in
321 * the pidns and the parent pid outside are identical. Using clone prevents
322 * this issue.
323 */
324 static void write_task_init_pid_exit(int sock, pid_t target)
325 {
326 char fnam[100];
327 pid_t pid;
328 int fd, ret;
329 size_t stack_size = sysconf(_SC_PAGESIZE);
330 void *stack = alloca(stack_size);
331
332 ret = snprintf(fnam, sizeof(fnam), "/proc/%d/ns/pid", (int)target);
333 if (ret < 0 || ret >= sizeof(fnam))
334 _exit(1);
335
336 fd = open(fnam, O_RDONLY);
337 if (fd < 0) {
338 perror("write_task_init_pid_exit open of ns/pid");
339 _exit(1);
340 }
341 if (setns(fd, 0)) {
342 perror("write_task_init_pid_exit setns 1");
343 close(fd);
344 _exit(1);
345 }
346 pid = clone(send_creds_clone_wrapper, stack + stack_size, SIGCHLD, &sock);
347 if (pid < 0)
348 _exit(1);
349 if (pid != 0) {
350 if (!wait_for_pid(pid))
351 _exit(1);
352 _exit(0);
353 }
354 }
355
356 static pid_t get_init_pid_for_task(pid_t task)
357 {
358 int sock[2];
359 pid_t pid;
360 pid_t ret = -1;
361 char v = '0';
362 struct ucred cred;
363
364 if (socketpair(AF_UNIX, SOCK_DGRAM, 0, sock) < 0) {
365 perror("socketpair");
366 return -1;
367 }
368
369 pid = fork();
370 if (pid < 0)
371 goto out;
372 if (!pid) {
373 close(sock[1]);
374 write_task_init_pid_exit(sock[0], task);
375 _exit(0);
376 }
377
378 if (!recv_creds(sock[1], &cred, &v))
379 goto out;
380 ret = cred.pid;
381
382 out:
383 close(sock[0]);
384 close(sock[1]);
385 if (pid > 0)
386 wait_for_pid(pid);
387 return ret;
388 }
389
390 #define LXCFS_PROC_PID_NS_LEN \
391 (STRLITERALLEN("/proc/") + INTTYPE_TO_STRLEN(uint64_t) + \
392 STRLITERALLEN("/ns/pid") + 1)
393
394 pid_t lookup_initpid_in_store(pid_t pid)
395 {
396 pid_t answer = 0;
397 char path[LXCFS_PROC_PID_NS_LEN];
398 struct stat st;
399 struct pidns_init_store *entry;
400
401 snprintf(path, sizeof(path), "/proc/%d/ns/pid", pid);
402
403 store_lock();
404 if (stat(path, &st))
405 goto out;
406
407 entry = lookup_verify_initpid(&st);
408 if (entry) {
409 answer = entry->initpid;
410 goto out;
411 }
412
413 answer = get_init_pid_for_task(pid);
414 if (answer > 0)
415 save_initpid(&st, answer);
416
417 out:
418 /*
419 * Prune at the end in case we're returning the value we were about to
420 * return.
421 */
422 prune_initpid_store();
423
424 store_unlock();
425
426 return answer;
427 }
428
429 /*
430 * Functions needed to setup cgroups in the __constructor__.
431 */
432
433 static bool umount_if_mounted(void)
434 {
435 if (umount2(BASEDIR, MNT_DETACH) < 0 && errno != EINVAL) {
436 lxcfs_error("Failed to unmount %s: %s.\n", BASEDIR, strerror(errno));
437 return false;
438 }
439 return true;
440 }
441
442 /* __typeof__ should be safe to use with all compilers. */
443 typedef __typeof__(((struct statfs *)NULL)->f_type) fs_type_magic;
444 static bool has_fs_type(const struct statfs *fs, fs_type_magic magic_val)
445 {
446 return (fs->f_type == (fs_type_magic)magic_val);
447 }
448
449 /*
450 * looking at fs/proc_namespace.c, it appears we can
451 * actually expect the rootfs entry to very specifically contain
452 * " - rootfs rootfs "
453 * IIUC, so long as we've chrooted so that rootfs is not our root,
454 * the rootfs entry should always be skipped in mountinfo contents.
455 */
456 static bool is_on_ramfs(void)
457 {
458 FILE *f;
459 char *p, *p2;
460 char *line = NULL;
461 size_t len = 0;
462 int i;
463
464 f = fopen("/proc/self/mountinfo", "r");
465 if (!f)
466 return false;
467
468 while (getline(&line, &len, f) != -1) {
469 for (p = line, i = 0; p && i < 4; i++)
470 p = strchr(p + 1, ' ');
471 if (!p)
472 continue;
473 p2 = strchr(p + 1, ' ');
474 if (!p2)
475 continue;
476 *p2 = '\0';
477 if (strcmp(p + 1, "/") == 0) {
478 // this is '/'. is it the ramfs?
479 p = strchr(p2 + 1, '-');
480 if (p && strncmp(p, "- rootfs rootfs ", 16) == 0) {
481 free(line);
482 fclose(f);
483 return true;
484 }
485 }
486 }
487 free(line);
488 fclose(f);
489 return false;
490 }
491
492 static int pivot_enter()
493 {
494 int ret = -1, oldroot = -1, newroot = -1;
495
496 oldroot = open("/", O_DIRECTORY | O_RDONLY);
497 if (oldroot < 0) {
498 lxcfs_error("%s\n", "Failed to open old root for fchdir.");
499 return ret;
500 }
501
502 newroot = open(ROOTDIR, O_DIRECTORY | O_RDONLY);
503 if (newroot < 0) {
504 lxcfs_error("%s\n", "Failed to open new root for fchdir.");
505 goto err;
506 }
507
508 /* change into new root fs */
509 if (fchdir(newroot) < 0) {
510 lxcfs_error("Failed to change directory to new rootfs: %s.\n", ROOTDIR);
511 goto err;
512 }
513
514 /* pivot_root into our new root fs */
515 if (pivot_root(".", ".") < 0) {
516 lxcfs_error("pivot_root() syscall failed: %s.\n", strerror(errno));
517 goto err;
518 }
519
520 /*
521 * At this point the old-root is mounted on top of our new-root.
522 * To unmounted it we must not be chdir'd into it, so escape back
523 * to the old-root.
524 */
525 if (fchdir(oldroot) < 0) {
526 lxcfs_error("%s\n", "Failed to enter old root.");
527 goto err;
528 }
529
530 if (umount2(".", MNT_DETACH) < 0) {
531 lxcfs_error("%s\n", "Failed to detach old root.");
532 goto err;
533 }
534
535 if (fchdir(newroot) < 0) {
536 lxcfs_error("%s\n", "Failed to re-enter new root.");
537 goto err;
538 }
539
540 ret = 0;
541
542 err:
543 if (oldroot > 0)
544 close(oldroot);
545 if (newroot > 0)
546 close(newroot);
547
548 return ret;
549 }
550
551 static int chroot_enter()
552 {
553 if (mount(ROOTDIR, "/", NULL, MS_REC | MS_BIND, NULL)) {
554 lxcfs_error("Failed to recursively bind-mount %s into /.", ROOTDIR);
555 return -1;
556 }
557
558 if (chroot(".") < 0) {
559 lxcfs_error("Call to chroot() failed: %s.\n", strerror(errno));
560 return -1;
561 }
562
563 if (chdir("/") < 0) {
564 lxcfs_error("Failed to change directory: %s.\n", strerror(errno));
565 return -1;
566 }
567
568 return 0;
569 }
570
571 static int permute_and_enter(void)
572 {
573 struct statfs sb;
574
575 if (statfs("/", &sb) < 0) {
576 lxcfs_error("%s\n", "Could not stat / mountpoint.");
577 return -1;
578 }
579
580 /* has_fs_type() is not reliable. When the ramfs is a tmpfs it will
581 * likely report TMPFS_MAGIC. Hence, when it reports no we still check
582 * /proc/1/mountinfo. */
583 if (has_fs_type(&sb, RAMFS_MAGIC) || is_on_ramfs())
584 return chroot_enter();
585
586 if (pivot_enter() < 0) {
587 lxcfs_error("%s\n", "Could not perform pivot root.");
588 return -1;
589 }
590
591 return 0;
592 }
593
594 /* Prepare our new clean root. */
595 static int permute_prepare(void)
596 {
597 if (mkdir(ROOTDIR, 0700) < 0 && errno != EEXIST) {
598 lxcfs_error("%s\n", "Failed to create directory for new root.");
599 return -1;
600 }
601
602 if (mount("/", ROOTDIR, NULL, MS_BIND, 0) < 0) {
603 lxcfs_error("Failed to bind-mount / for new root: %s.\n", strerror(errno));
604 return -1;
605 }
606
607 if (mount(RUNTIME_PATH, ROOTDIR RUNTIME_PATH, NULL, MS_BIND, 0) < 0) {
608 lxcfs_error("Failed to bind-mount /run into new root: %s.\n", strerror(errno));
609 return -1;
610 }
611
612 if (mount(BASEDIR, ROOTDIR BASEDIR, NULL, MS_REC | MS_MOVE, 0) < 0) {
613 printf("Failed to move " BASEDIR " into new root: %s.\n", strerror(errno));
614 return -1;
615 }
616
617 return 0;
618 }
619
620 /* Calls chroot() on ramfs, pivot_root() in all other cases. */
621 static bool permute_root(void)
622 {
623 /* Prepare new root. */
624 if (permute_prepare() < 0)
625 return false;
626
627 /* Pivot into new root. */
628 if (permute_and_enter() < 0)
629 return false;
630
631 return true;
632 }
633
634 static bool cgfs_prepare_mounts(void)
635 {
636 if (!mkdir_p(BASEDIR, 0700)) {
637 lxcfs_error("%s\n", "Failed to create lxcfs cgroup mountpoint.");
638 return false;
639 }
640
641 if (!umount_if_mounted()) {
642 lxcfs_error("%s\n", "Failed to clean up old lxcfs cgroup mountpoint.");
643 return false;
644 }
645
646 if (unshare(CLONE_NEWNS) < 0) {
647 lxcfs_error("Failed to unshare mount namespace: %s.\n", strerror(errno));
648 return false;
649 }
650
651 cgroup_ops->mntns_fd = preserve_ns(getpid(), "mnt");
652 if (cgroup_ops->mntns_fd < 0) {
653 lxcfs_error("Failed to preserve mount namespace: %s.\n", strerror(errno));
654 return false;
655 }
656
657 if (mount(NULL, "/", NULL, MS_REC | MS_PRIVATE, 0) < 0) {
658 lxcfs_error("Failed to remount / private: %s.\n", strerror(errno));
659 return false;
660 }
661
662 if (mount("tmpfs", BASEDIR, "tmpfs", 0, "size=100000,mode=700") < 0) {
663 lxcfs_error("%s\n", "Failed to mount tmpfs over lxcfs cgroup mountpoint.");
664 return false;
665 }
666
667 return true;
668 }
669
670 static bool cgfs_mount_hierarchies(void)
671 {
672 if (!mkdir_p(BASEDIR DEFAULT_CGROUP_MOUNTPOINT, 0755))
673 return false;
674
675 if (!cgroup_ops->mount(cgroup_ops, BASEDIR))
676 return false;
677
678 for (struct hierarchy **h = cgroup_ops->hierarchies; h && *h; h++) {
679 __do_free char *path = must_make_path(BASEDIR, (*h)->mountpoint, NULL);
680 (*h)->fd = open(path, O_DIRECTORY | O_CLOEXEC | O_NOFOLLOW);
681 if ((*h)->fd < 0)
682 return false;
683 }
684
685 return true;
686 }
687
688 static bool cgfs_setup_controllers(void)
689 {
690 if (!cgfs_prepare_mounts())
691 return false;
692
693 if (!cgfs_mount_hierarchies()) {
694 lxcfs_error("%s\n", "Failed to set up private lxcfs cgroup mounts.");
695 return false;
696 }
697
698 if (!permute_root())
699 return false;
700
701 return true;
702 }
703
704 static void __attribute__((constructor)) lxcfs_init(void)
705 {
706 __do_close_prot_errno int init_ns = -EBADF, pidfd = -EBADF;
707 int i = 0;
708 pid_t pid;
709 char *cret;
710 char cwd[MAXPATHLEN];
711
712 cgroup_ops = cgroup_init();
713 if (!cgroup_ops)
714 log_exit("Failed to initialize cgroup support");
715
716 /* Preserve initial namespace. */
717 pid = getpid();
718 init_ns = preserve_ns(pid, "mnt");
719 if (init_ns < 0)
720 log_exit("Failed to preserve initial mount namespace");
721
722 cret = getcwd(cwd, MAXPATHLEN);
723 if (!cret)
724 log_exit("%s - Could not retrieve current working directory", strerror(errno));
725
726 /* This function calls unshare(CLONE_NEWNS) our initial mount namespace
727 * to privately mount lxcfs cgroups. */
728 if (!cgfs_setup_controllers())
729 log_exit("Failed to setup private cgroup mounts for lxcfs");
730
731 if (setns(init_ns, 0) < 0)
732 log_exit("%s - Failed to switch back to initial mount namespace", strerror(errno));
733
734 if (!cret || chdir(cwd) < 0)
735 log_exit("%s - Could not change back to original working directory", strerror(errno));
736
737 if (!init_cpuview())
738 log_exit("Failed to init CPU view");
739
740 fprintf(stderr, "mount namespace: %d\n", cgroup_ops->mntns_fd);
741 fprintf(stderr, "hierarchies:\n");
742
743 for (struct hierarchy **h = cgroup_ops->hierarchies; h && *h; h++, i++) {
744 __do_free char *controllers = lxc_string_join(",", (const char **)(*h)->controllers, false);
745 fprintf(stderr, " %2d: fd: %3d: %s\n", i, (*h)->fd, controllers ?: "");
746 }
747
748 pidfd = pidfd_open(pid, 0);
749 if (pidfd >= 0 && pidfd_send_signal(pidfd, 0, NULL, 0) == 0) {
750 can_use_pidfd = true;
751 lxcfs_error("Kernel supports pidfds");
752 }
753 }
754
755 static void __attribute__((destructor)) lxcfs_exit(void)
756 {
757 lxcfs_debug("%s\n", "Running destructor for liblxcfs");
758 free_cpuview();
759 cgroup_exit(cgroup_ops);
760 }