]> git.proxmox.com Git - mirror_lxcfs.git/blob - bindings.c
Merge pull request #329 from brauner/master
[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 *entry = 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 entry = malloc(sizeof(*entry));
256 if (entry)
257 return;
258
259 ino_hash = HASH(entry->ino);
260 *entry = (struct pidns_init_store){
261 .ino = sb->st_ino,
262 .initpid = pid,
263 .ctime = st.st_ctime,
264 .next = pidns_hash_table[ino_hash],
265 .lastcheck = time(NULL),
266 .init_pidfd = move_fd(pidfd),
267 };
268 pidns_hash_table[ino_hash] = move_ptr(entry);
269
270 lxcfs_debug("Added cache entry %d for pid %d to init pid cache", ino_hash, pid);
271 }
272
273 /*
274 * Given the stat(2) info for a nsfd pid inode, lookup the init_pid_store
275 * entry for the inode number and creation time. Verify that the init pid
276 * is still valid. If not, remove it. Return the entry if valid, NULL
277 * otherwise.
278 * Must be called under store_lock
279 */
280 static struct pidns_init_store *lookup_verify_initpid(struct stat *sb)
281 {
282 struct pidns_init_store *entry = pidns_hash_table[HASH(sb->st_ino)];
283
284 while (entry) {
285 if (entry->ino == sb->st_ino) {
286 if (initpid_still_valid(entry)) {
287 entry->lastcheck = time(NULL);
288 return entry;
289 }
290
291 remove_initpid(entry);
292 return NULL;
293 }
294 entry = entry->next;
295 }
296
297 return NULL;
298 }
299
300 static int send_creds_clone_wrapper(void *arg)
301 {
302 struct ucred cred;
303 char v;
304 int sock = *(int *)arg;
305
306 /* we are the child */
307 cred.uid = 0;
308 cred.gid = 0;
309 cred.pid = 1;
310 v = '1';
311 if (send_creds(sock, &cred, v, true) != SEND_CREDS_OK)
312 return 1;
313 return 0;
314 }
315
316 /*
317 * clone a task which switches to @task's namespace and writes '1'.
318 * over a unix sock so we can read the task's reaper's pid in our
319 * namespace
320 *
321 * Note: glibc's fork() does not respect pidns, which can lead to failed
322 * assertions inside glibc (and thus failed forks) if the child's pid in
323 * the pidns and the parent pid outside are identical. Using clone prevents
324 * this issue.
325 */
326 static void write_task_init_pid_exit(int sock, pid_t target)
327 {
328 char fnam[100];
329 pid_t pid;
330 int fd, ret;
331 size_t stack_size = sysconf(_SC_PAGESIZE);
332 void *stack = alloca(stack_size);
333
334 ret = snprintf(fnam, sizeof(fnam), "/proc/%d/ns/pid", (int)target);
335 if (ret < 0 || ret >= sizeof(fnam))
336 _exit(1);
337
338 fd = open(fnam, O_RDONLY);
339 if (fd < 0) {
340 perror("write_task_init_pid_exit open of ns/pid");
341 _exit(1);
342 }
343 if (setns(fd, 0)) {
344 perror("write_task_init_pid_exit setns 1");
345 close(fd);
346 _exit(1);
347 }
348 pid = clone(send_creds_clone_wrapper, stack + stack_size, SIGCHLD, &sock);
349 if (pid < 0)
350 _exit(1);
351 if (pid != 0) {
352 if (!wait_for_pid(pid))
353 _exit(1);
354 _exit(0);
355 }
356 }
357
358 static pid_t get_init_pid_for_task(pid_t task)
359 {
360 int sock[2];
361 pid_t pid;
362 pid_t ret = -1;
363 char v = '0';
364 struct ucred cred;
365
366 if (socketpair(AF_UNIX, SOCK_DGRAM, 0, sock) < 0) {
367 perror("socketpair");
368 return -1;
369 }
370
371 pid = fork();
372 if (pid < 0)
373 goto out;
374 if (!pid) {
375 close(sock[1]);
376 write_task_init_pid_exit(sock[0], task);
377 _exit(0);
378 }
379
380 if (!recv_creds(sock[1], &cred, &v))
381 goto out;
382 ret = cred.pid;
383
384 out:
385 close(sock[0]);
386 close(sock[1]);
387 if (pid > 0)
388 wait_for_pid(pid);
389 return ret;
390 }
391
392 #define LXCFS_PROC_PID_NS_LEN \
393 (STRLITERALLEN("/proc/") + INTTYPE_TO_STRLEN(uint64_t) + \
394 STRLITERALLEN("/ns/pid") + 1)
395
396 pid_t lookup_initpid_in_store(pid_t pid)
397 {
398 pid_t answer = 0;
399 char path[LXCFS_PROC_PID_NS_LEN];
400 struct stat st;
401 struct pidns_init_store *entry;
402
403 snprintf(path, sizeof(path), "/proc/%d/ns/pid", pid);
404
405 store_lock();
406 if (stat(path, &st))
407 goto out;
408
409 entry = lookup_verify_initpid(&st);
410 if (entry) {
411 answer = entry->initpid;
412 goto out;
413 }
414
415 answer = get_init_pid_for_task(pid);
416 if (answer > 0)
417 save_initpid(&st, answer);
418
419 out:
420 /*
421 * Prune at the end in case we're returning the value we were about to
422 * return.
423 */
424 prune_initpid_store();
425
426 store_unlock();
427
428 return answer;
429 }
430
431 /*
432 * Functions needed to setup cgroups in the __constructor__.
433 */
434
435 static bool umount_if_mounted(void)
436 {
437 if (umount2(BASEDIR, MNT_DETACH) < 0 && errno != EINVAL) {
438 lxcfs_error("Failed to unmount %s: %s.\n", BASEDIR, strerror(errno));
439 return false;
440 }
441 return true;
442 }
443
444 /* __typeof__ should be safe to use with all compilers. */
445 typedef __typeof__(((struct statfs *)NULL)->f_type) fs_type_magic;
446 static bool has_fs_type(const struct statfs *fs, fs_type_magic magic_val)
447 {
448 return (fs->f_type == (fs_type_magic)magic_val);
449 }
450
451 /*
452 * looking at fs/proc_namespace.c, it appears we can
453 * actually expect the rootfs entry to very specifically contain
454 * " - rootfs rootfs "
455 * IIUC, so long as we've chrooted so that rootfs is not our root,
456 * the rootfs entry should always be skipped in mountinfo contents.
457 */
458 static bool is_on_ramfs(void)
459 {
460 FILE *f;
461 char *p, *p2;
462 char *line = NULL;
463 size_t len = 0;
464 int i;
465
466 f = fopen("/proc/self/mountinfo", "r");
467 if (!f)
468 return false;
469
470 while (getline(&line, &len, f) != -1) {
471 for (p = line, i = 0; p && i < 4; i++)
472 p = strchr(p + 1, ' ');
473 if (!p)
474 continue;
475 p2 = strchr(p + 1, ' ');
476 if (!p2)
477 continue;
478 *p2 = '\0';
479 if (strcmp(p + 1, "/") == 0) {
480 // this is '/'. is it the ramfs?
481 p = strchr(p2 + 1, '-');
482 if (p && strncmp(p, "- rootfs rootfs ", 16) == 0) {
483 free(line);
484 fclose(f);
485 return true;
486 }
487 }
488 }
489 free(line);
490 fclose(f);
491 return false;
492 }
493
494 static int pivot_enter()
495 {
496 int ret = -1, oldroot = -1, newroot = -1;
497
498 oldroot = open("/", O_DIRECTORY | O_RDONLY);
499 if (oldroot < 0) {
500 lxcfs_error("%s\n", "Failed to open old root for fchdir.");
501 return ret;
502 }
503
504 newroot = open(ROOTDIR, O_DIRECTORY | O_RDONLY);
505 if (newroot < 0) {
506 lxcfs_error("%s\n", "Failed to open new root for fchdir.");
507 goto err;
508 }
509
510 /* change into new root fs */
511 if (fchdir(newroot) < 0) {
512 lxcfs_error("Failed to change directory to new rootfs: %s.\n", ROOTDIR);
513 goto err;
514 }
515
516 /* pivot_root into our new root fs */
517 if (pivot_root(".", ".") < 0) {
518 lxcfs_error("pivot_root() syscall failed: %s.\n", strerror(errno));
519 goto err;
520 }
521
522 /*
523 * At this point the old-root is mounted on top of our new-root.
524 * To unmounted it we must not be chdir'd into it, so escape back
525 * to the old-root.
526 */
527 if (fchdir(oldroot) < 0) {
528 lxcfs_error("%s\n", "Failed to enter old root.");
529 goto err;
530 }
531
532 if (umount2(".", MNT_DETACH) < 0) {
533 lxcfs_error("%s\n", "Failed to detach old root.");
534 goto err;
535 }
536
537 if (fchdir(newroot) < 0) {
538 lxcfs_error("%s\n", "Failed to re-enter new root.");
539 goto err;
540 }
541
542 ret = 0;
543
544 err:
545 if (oldroot > 0)
546 close(oldroot);
547 if (newroot > 0)
548 close(newroot);
549
550 return ret;
551 }
552
553 static int chroot_enter()
554 {
555 if (mount(ROOTDIR, "/", NULL, MS_REC | MS_BIND, NULL)) {
556 lxcfs_error("Failed to recursively bind-mount %s into /.", ROOTDIR);
557 return -1;
558 }
559
560 if (chroot(".") < 0) {
561 lxcfs_error("Call to chroot() failed: %s.\n", strerror(errno));
562 return -1;
563 }
564
565 if (chdir("/") < 0) {
566 lxcfs_error("Failed to change directory: %s.\n", strerror(errno));
567 return -1;
568 }
569
570 return 0;
571 }
572
573 static int permute_and_enter(void)
574 {
575 struct statfs sb;
576
577 if (statfs("/", &sb) < 0) {
578 lxcfs_error("%s\n", "Could not stat / mountpoint.");
579 return -1;
580 }
581
582 /* has_fs_type() is not reliable. When the ramfs is a tmpfs it will
583 * likely report TMPFS_MAGIC. Hence, when it reports no we still check
584 * /proc/1/mountinfo. */
585 if (has_fs_type(&sb, RAMFS_MAGIC) || is_on_ramfs())
586 return chroot_enter();
587
588 if (pivot_enter() < 0) {
589 lxcfs_error("%s\n", "Could not perform pivot root.");
590 return -1;
591 }
592
593 return 0;
594 }
595
596 /* Prepare our new clean root. */
597 static int permute_prepare(void)
598 {
599 if (mkdir(ROOTDIR, 0700) < 0 && errno != EEXIST) {
600 lxcfs_error("%s\n", "Failed to create directory for new root.");
601 return -1;
602 }
603
604 if (mount("/", ROOTDIR, NULL, MS_BIND, 0) < 0) {
605 lxcfs_error("Failed to bind-mount / for new root: %s.\n", strerror(errno));
606 return -1;
607 }
608
609 if (mount(RUNTIME_PATH, ROOTDIR RUNTIME_PATH, NULL, MS_BIND, 0) < 0) {
610 lxcfs_error("Failed to bind-mount /run into new root: %s.\n", strerror(errno));
611 return -1;
612 }
613
614 if (mount(BASEDIR, ROOTDIR BASEDIR, NULL, MS_REC | MS_MOVE, 0) < 0) {
615 printf("Failed to move " BASEDIR " into new root: %s.\n", strerror(errno));
616 return -1;
617 }
618
619 return 0;
620 }
621
622 /* Calls chroot() on ramfs, pivot_root() in all other cases. */
623 static bool permute_root(void)
624 {
625 /* Prepare new root. */
626 if (permute_prepare() < 0)
627 return false;
628
629 /* Pivot into new root. */
630 if (permute_and_enter() < 0)
631 return false;
632
633 return true;
634 }
635
636 static bool cgfs_prepare_mounts(void)
637 {
638 if (!mkdir_p(BASEDIR, 0700)) {
639 lxcfs_error("%s\n", "Failed to create lxcfs cgroup mountpoint.");
640 return false;
641 }
642
643 if (!umount_if_mounted()) {
644 lxcfs_error("%s\n", "Failed to clean up old lxcfs cgroup mountpoint.");
645 return false;
646 }
647
648 if (unshare(CLONE_NEWNS) < 0) {
649 lxcfs_error("Failed to unshare mount namespace: %s.\n", strerror(errno));
650 return false;
651 }
652
653 cgroup_ops->mntns_fd = preserve_ns(getpid(), "mnt");
654 if (cgroup_ops->mntns_fd < 0) {
655 lxcfs_error("Failed to preserve mount namespace: %s.\n", strerror(errno));
656 return false;
657 }
658
659 if (mount(NULL, "/", NULL, MS_REC | MS_PRIVATE, 0) < 0) {
660 lxcfs_error("Failed to remount / private: %s.\n", strerror(errno));
661 return false;
662 }
663
664 if (mount("tmpfs", BASEDIR, "tmpfs", 0, "size=100000,mode=700") < 0) {
665 lxcfs_error("%s\n", "Failed to mount tmpfs over lxcfs cgroup mountpoint.");
666 return false;
667 }
668
669 return true;
670 }
671
672 static bool cgfs_mount_hierarchies(void)
673 {
674 if (!mkdir_p(BASEDIR DEFAULT_CGROUP_MOUNTPOINT, 0755))
675 return false;
676
677 if (!cgroup_ops->mount(cgroup_ops, BASEDIR))
678 return false;
679
680 for (struct hierarchy **h = cgroup_ops->hierarchies; h && *h; h++) {
681 __do_free char *path = must_make_path(BASEDIR, (*h)->mountpoint, NULL);
682 (*h)->fd = open(path, O_DIRECTORY | O_CLOEXEC | O_NOFOLLOW);
683 if ((*h)->fd < 0)
684 return false;
685 }
686
687 return true;
688 }
689
690 static bool cgfs_setup_controllers(void)
691 {
692 if (!cgfs_prepare_mounts())
693 return false;
694
695 if (!cgfs_mount_hierarchies()) {
696 lxcfs_error("%s\n", "Failed to set up private lxcfs cgroup mounts.");
697 return false;
698 }
699
700 if (!permute_root())
701 return false;
702
703 return true;
704 }
705
706 static void __attribute__((constructor)) lxcfs_init(void)
707 {
708 __do_close_prot_errno int init_ns = -EBADF, pidfd = -EBADF;
709 int i = 0;
710 pid_t pid;
711 char *cret;
712 char cwd[MAXPATHLEN];
713
714 cgroup_ops = cgroup_init();
715 if (!cgroup_ops)
716 log_exit("Failed to initialize cgroup support");
717
718 /* Preserve initial namespace. */
719 pid = getpid();
720 init_ns = preserve_ns(pid, "mnt");
721 if (init_ns < 0)
722 log_exit("Failed to preserve initial mount namespace");
723
724 cret = getcwd(cwd, MAXPATHLEN);
725 if (!cret)
726 log_exit("%s - Could not retrieve current working directory", strerror(errno));
727
728 /* This function calls unshare(CLONE_NEWNS) our initial mount namespace
729 * to privately mount lxcfs cgroups. */
730 if (!cgfs_setup_controllers())
731 log_exit("Failed to setup private cgroup mounts for lxcfs");
732
733 if (setns(init_ns, 0) < 0)
734 log_exit("%s - Failed to switch back to initial mount namespace", strerror(errno));
735
736 if (!cret || chdir(cwd) < 0)
737 log_exit("%s - Could not change back to original working directory", strerror(errno));
738
739 if (!init_cpuview())
740 log_exit("Failed to init CPU view");
741
742 fprintf(stderr, "mount namespace: %d\n", cgroup_ops->mntns_fd);
743 fprintf(stderr, "hierarchies:\n");
744
745 for (struct hierarchy **h = cgroup_ops->hierarchies; h && *h; h++, i++) {
746 __do_free char *controllers = lxc_string_join(",", (const char **)(*h)->controllers, false);
747 fprintf(stderr, " %2d: fd: %3d: %s\n", i, (*h)->fd, controllers ?: "");
748 }
749
750 pidfd = pidfd_open(pid, 0);
751 if (pidfd >= 0 && pidfd_send_signal(pidfd, 0, NULL, 0) == 0) {
752 can_use_pidfd = true;
753 lxcfs_error("Kernel supports pidfds");
754 }
755 }
756
757 static void __attribute__((destructor)) lxcfs_exit(void)
758 {
759 lxcfs_debug("%s\n", "Running destructor for liblxcfs");
760 free_cpuview();
761 cgroup_exit(cgroup_ops);
762 }