3 * Copyright © 2014-2016 Canonical, Inc
4 * Author: Serge Hallyn <serge.hallyn@ubuntu.com>
6 * See COPYING file for details.
13 #ifndef FUSE_USE_VERSION
14 #define FUSE_USE_VERSION 26
17 #define _FILE_OFFSET_BITS 64
36 #include <linux/magic.h>
37 #include <linux/sched.h>
38 #include <sys/epoll.h>
40 #include <sys/mount.h>
41 #include <sys/param.h>
43 #include <sys/socket.h>
44 #include <sys/syscall.h>
45 #include <sys/sysinfo.h>
48 #include "api_extensions.h"
51 #include "cgroup_fuse.h"
52 #include "cgroups/cgroup.h"
53 #include "cgroups/cgroup_utils.h"
54 #include "memory_utils.h"
55 #include "proc_cpuview.h"
58 static bool can_use_pidfd
;
60 /* Define pivot_root() if missing from the C library */
61 #ifndef HAVE_PIVOT_ROOT
62 static int pivot_root(const char *new_root
, const char *put_old
)
64 #ifdef __NR_pivot_root
65 return syscall(__NR_pivot_root
, new_root
, put_old
);
72 extern int pivot_root(const char *new_root
, const char *put_old
);
76 * A table caching which pid is init for a pid namespace.
77 * When looking up which pid is init for $qpid, we first
78 * 1. Stat /proc/$qpid/ns/pid.
79 * 2. Check whether the ino_t is in our store.
80 * a. if not, fork a child in qpid's ns to send us
81 * ucred.pid = 1, and read the initpid. Cache
82 * initpid and creation time for /proc/initpid
83 * in a new store entry.
84 * b. if so, verify that /proc/initpid still matches
85 * what we have saved. If not, clear the store
86 * entry and go back to a. If so, return the
89 struct pidns_init_store
{
90 ino_t ino
; /* inode number for /proc/$pid/ns/pid */
91 pid_t initpid
; /* the pid of nit in that ns */
93 long int ctime
; /* the time at which /proc/$initpid was created */
94 struct pidns_init_store
*next
;
98 /* lol - look at how they are allocated in the kernel */
99 #define PIDNS_HASH_SIZE 4096
100 #define HASH(x) ((x) % PIDNS_HASH_SIZE)
102 static struct pidns_init_store
*pidns_hash_table
[PIDNS_HASH_SIZE
];
103 static pthread_mutex_t pidns_store_mutex
= PTHREAD_MUTEX_INITIALIZER
;
105 static void lock_mutex(pthread_mutex_t
*l
)
109 ret
= pthread_mutex_lock(l
);
111 log_exit("%s - returned %d\n", strerror(ret
), ret
);
114 struct cgroup_ops
*cgroup_ops
;
116 static void unlock_mutex(pthread_mutex_t
*l
)
120 ret
= pthread_mutex_unlock(l
);
122 log_exit("%s - returned %d\n", strerror(ret
), ret
);
125 static void store_lock(void)
127 lock_mutex(&pidns_store_mutex
);
130 static void store_unlock(void)
132 unlock_mutex(&pidns_store_mutex
);
137 * <pid-as-str> = INTTYPE_TO_STRLEN(pid_t)
141 #define LXCFS_PROC_PID_LEN \
142 (STRLITERALLEN("/proc/") + INTTYPE_TO_STRLEN(uint64_t) + +1)
144 /* Must be called under store_lock */
145 static bool initpid_still_valid(struct pidns_init_store
*entry
)
149 if (entry
->init_pidfd
>= 0) {
150 if (pidfd_send_signal(entry
->init_pidfd
, 0, NULL
, 0))
154 char path
[LXCFS_PROC_PID_LEN
];
156 snprintf(path
, sizeof(path
), "/proc/%d", entry
->initpid
);
158 if (stat(path
, &st
) || entry
->ctime
!= st
.st_ctime
)
165 /* Must be called under store_lock */
166 static void remove_initpid(struct pidns_init_store
*entry
)
168 struct pidns_init_store
*it
;
171 lxcfs_debug("Removing cached entry for pid %d from init pid cache",
174 ino_hash
= HASH(entry
->ino
);
175 if (pidns_hash_table
[ino_hash
] == entry
) {
176 pidns_hash_table
[ino_hash
] = entry
->next
;
177 close_prot_errno_disarm(entry
->init_pidfd
);
182 it
= pidns_hash_table
[ino_hash
];
184 if (it
->next
== entry
) {
185 it
->next
= entry
->next
;
186 close_prot_errno_disarm(entry
->init_pidfd
);
195 /* Must be called under store_lock */
196 static void prune_initpid_store(void)
198 static long int last_prune
= 0;
199 long int now
, threshold
;
202 last_prune
= time(NULL
);
207 if (now
< last_prune
+ PURGE_SECS
)
210 lxcfs_debug("Pruning init pid cache");
213 threshold
= now
- 2 * PURGE_SECS
;
215 for (int i
= 0; i
< PIDNS_HASH_SIZE
; i
++) {
216 for (struct pidns_init_store
*entry
= pidns_hash_table
[i
], *prev
= NULL
; entry
;) {
217 if (entry
->lastcheck
< threshold
) {
218 struct pidns_init_store
*cur
= entry
;
220 lxcfs_debug("Removed cache entry for pid %d to init pid cache", cur
->initpid
);
223 prev
->next
= entry
->next
;
225 pidns_hash_table
[i
] = entry
->next
;
227 close_prot_errno_disarm(cur
->init_pidfd
);
237 /* Must be called under store_lock */
238 static void save_initpid(struct stat
*sb
, pid_t pid
)
240 __do_free
struct pidns_init_store
*entry
= NULL
;
241 __do_close_prot_errno
int pidfd
= -EBADF
;
242 char path
[LXCFS_PROC_PID_LEN
];
243 struct lxcfs_opts
*opts
= fuse_get_context()->private_data
;
247 if (opts
&& opts
->use_pidfd
&& can_use_pidfd
) {
248 pidfd
= pidfd_open(pid
, 0);
253 snprintf(path
, sizeof(path
), "/proc/%d", pid
);
257 entry
= malloc(sizeof(*entry
));
261 ino_hash
= HASH(entry
->ino
);
262 *entry
= (struct pidns_init_store
){
265 .ctime
= st
.st_ctime
,
266 .next
= pidns_hash_table
[ino_hash
],
267 .lastcheck
= time(NULL
),
268 .init_pidfd
= move_fd(pidfd
),
270 pidns_hash_table
[ino_hash
] = move_ptr(entry
);
272 lxcfs_debug("Added cache entry %d for pid %d to init pid cache", ino_hash
, pid
);
276 * Given the stat(2) info for a nsfd pid inode, lookup the init_pid_store
277 * entry for the inode number and creation time. Verify that the init pid
278 * is still valid. If not, remove it. Return the entry if valid, NULL
280 * Must be called under store_lock
282 static struct pidns_init_store
*lookup_verify_initpid(struct stat
*sb
)
284 struct pidns_init_store
*entry
= pidns_hash_table
[HASH(sb
->st_ino
)];
287 if (entry
->ino
== sb
->st_ino
) {
288 if (initpid_still_valid(entry
)) {
289 entry
->lastcheck
= time(NULL
);
293 remove_initpid(entry
);
302 static int send_creds_clone_wrapper(void *arg
)
304 int sock
= PTR_TO_INT(arg
);
305 char v
= '1'; /* we are the child */
306 struct ucred cred
= {
312 return send_creds(sock
, &cred
, v
, true) != SEND_CREDS_OK
;
316 * Let's use the "standard stack limit" (i.e. glibc thread size default) for
319 #define __LXCFS_STACK_SIZE (8 * 1024 * 1024)
320 static pid_t
lxcfs_clone(int (*fn
)(void *), void *arg
, int flags
)
325 stack
= malloc(__LXCFS_STACK_SIZE
);
327 return ret_errno(ENOMEM
);
330 ret
= __clone2(fn
, stack
, __LXCFS_STACK_SIZE
, flags
| SIGCHLD
, arg
, NULL
);
332 ret
= clone(fn
, stack
+ __LXCFS_STACK_SIZE
, flags
| SIGCHLD
, arg
, NULL
);
337 #define LXCFS_PROC_PID_NS_LEN \
338 (STRLITERALLEN("/proc/") + INTTYPE_TO_STRLEN(uint64_t) + \
339 STRLITERALLEN("/ns/pid") + 1)
342 * clone a task which switches to @task's namespace and writes '1'.
343 * over a unix sock so we can read the task's reaper's pid in our
346 * Note: glibc's fork() does not respect pidns, which can lead to failed
347 * assertions inside glibc (and thus failed forks) if the child's pid in
348 * the pidns and the parent pid outside are identical. Using clone prevents
351 static void write_task_init_pid_exit(int sock
, pid_t target
)
353 __do_close_prot_errno
int fd
= -EBADF
;
354 char path
[LXCFS_PROC_PID_NS_LEN
];
357 snprintf(path
, sizeof(path
), "/proc/%d/ns/pid", (int)target
);
358 fd
= open(path
, O_RDONLY
| O_CLOEXEC
);
360 log_exit("write_task_init_pid_exit open of ns/pid");
363 log_exit("Failed to setns to pid namespace of process %d", target
);
365 pid
= lxcfs_clone(send_creds_clone_wrapper
, INT_TO_PTR(sock
), 0);
370 if (!wait_for_pid(pid
))
377 static pid_t
get_init_pid_for_task(pid_t task
)
385 if (socketpair(AF_UNIX
, SOCK_DGRAM
, 0, sock
) < 0)
394 write_task_init_pid_exit(sock
[0], task
);
398 if (!recv_creds(sock
[1], &cred
, &v
))
412 pid_t
lookup_initpid_in_store(pid_t pid
)
415 char path
[LXCFS_PROC_PID_NS_LEN
];
417 struct pidns_init_store
*entry
;
419 snprintf(path
, sizeof(path
), "/proc/%d/ns/pid", pid
);
425 entry
= lookup_verify_initpid(&st
);
427 answer
= entry
->initpid
;
431 answer
= get_init_pid_for_task(pid
);
433 save_initpid(&st
, answer
);
437 * Prune at the end in case we're returning the value we were about to
440 prune_initpid_store();
448 * Functions needed to setup cgroups in the __constructor__.
451 static bool umount_if_mounted(void)
453 if (umount2(BASEDIR
, MNT_DETACH
) < 0 && errno
!= EINVAL
) {
454 lxcfs_error("Failed to unmount %s: %s.\n", BASEDIR
, strerror(errno
));
460 /* __typeof__ should be safe to use with all compilers. */
461 typedef __typeof__(((struct statfs
*)NULL
)->f_type
) fs_type_magic
;
462 static bool has_fs_type(const struct statfs
*fs
, fs_type_magic magic_val
)
464 return (fs
->f_type
== (fs_type_magic
)magic_val
);
468 * looking at fs/proc_namespace.c, it appears we can
469 * actually expect the rootfs entry to very specifically contain
470 * " - rootfs rootfs "
471 * IIUC, so long as we've chrooted so that rootfs is not our root,
472 * the rootfs entry should always be skipped in mountinfo contents.
474 static bool is_on_ramfs(void)
476 __do_free
char *line
= NULL
;
477 __do_free
void *fopen_cache
= NULL
;
478 __do_fclose
FILE *f
= NULL
;
481 f
= fopen_cached("/proc/self/mountinfo", "re", &fopen_cache
);
485 while (getline(&line
, &len
, f
) != -1) {
489 for (p
= line
, i
= 0; p
&& i
< 4; i
++)
490 p
= strchr(p
+ 1, ' ');
494 p2
= strchr(p
+ 1, ' ');
498 if (strcmp(p
+ 1, "/") == 0) {
499 /* This is '/'. Is it the ramfs? */
500 p
= strchr(p2
+ 1, '-');
501 if (p
&& strncmp(p
, "- rootfs rootfs ", 16) == 0)
509 static int pivot_enter()
511 __do_close_prot_errno
int oldroot
= -EBADF
, newroot
= -EBADF
;
513 oldroot
= open("/", O_DIRECTORY
| O_RDONLY
);
515 return log_error_errno(-1, errno
,
516 "Failed to open old root for fchdir");
518 newroot
= open(ROOTDIR
, O_DIRECTORY
| O_RDONLY
);
520 return log_error_errno(-1, errno
,
521 "Failed to open new root for fchdir");
523 /* change into new root fs */
524 if (fchdir(newroot
) < 0)
525 return log_error_errno(-1,
526 errno
, "Failed to change directory to new rootfs: %s",
529 /* pivot_root into our new root fs */
530 if (pivot_root(".", ".") < 0)
531 return log_error_errno(-1, errno
,
532 "pivot_root() syscall failed: %s",
536 * At this point the old-root is mounted on top of our new-root.
537 * To unmounted it we must not be chdir'd into it, so escape back
540 if (fchdir(oldroot
) < 0)
541 return log_error_errno(-1, errno
, "Failed to enter old root");
543 if (umount2(".", MNT_DETACH
) < 0)
544 return log_error_errno(-1, errno
, "Failed to detach old root");
546 if (fchdir(newroot
) < 0)
547 return log_error_errno(-1, errno
, "Failed to re-enter new root");
552 static int chroot_enter()
554 if (mount(ROOTDIR
, "/", NULL
, MS_REC
| MS_BIND
, NULL
)) {
555 lxcfs_error("Failed to recursively bind-mount %s into /.", ROOTDIR
);
559 if (chroot(".") < 0) {
560 lxcfs_error("Call to chroot() failed: %s.\n", strerror(errno
));
564 if (chdir("/") < 0) {
565 lxcfs_error("Failed to change directory: %s.\n", strerror(errno
));
572 static int permute_and_enter(void)
576 if (statfs("/", &sb
) < 0) {
577 lxcfs_error("%s\n", "Could not stat / mountpoint.");
581 /* has_fs_type() is not reliable. When the ramfs is a tmpfs it will
582 * likely report TMPFS_MAGIC. Hence, when it reports no we still check
583 * /proc/1/mountinfo. */
584 if (has_fs_type(&sb
, RAMFS_MAGIC
) || is_on_ramfs())
585 return chroot_enter();
587 if (pivot_enter() < 0) {
588 lxcfs_error("%s\n", "Could not perform pivot root.");
595 /* Prepare our new clean root. */
596 static int permute_prepare(void)
598 if (mkdir(ROOTDIR
, 0700) < 0 && errno
!= EEXIST
) {
599 lxcfs_error("%s\n", "Failed to create directory for new root.");
603 if (mount("/", ROOTDIR
, NULL
, MS_BIND
, 0) < 0) {
604 lxcfs_error("Failed to bind-mount / for new root: %s.\n", strerror(errno
));
608 if (mount(RUNTIME_PATH
, ROOTDIR RUNTIME_PATH
, NULL
, MS_BIND
, 0) < 0) {
609 lxcfs_error("Failed to bind-mount /run into new root: %s.\n", strerror(errno
));
613 if (mount(BASEDIR
, ROOTDIR BASEDIR
, NULL
, MS_REC
| MS_MOVE
, 0) < 0) {
614 printf("Failed to move " BASEDIR
" into new root: %s.\n", strerror(errno
));
621 /* Calls chroot() on ramfs, pivot_root() in all other cases. */
622 static bool permute_root(void)
624 /* Prepare new root. */
625 if (permute_prepare() < 0)
628 /* Pivot into new root. */
629 if (permute_and_enter() < 0)
635 static bool cgfs_prepare_mounts(void)
637 if (!mkdir_p(BASEDIR
, 0700)) {
638 lxcfs_error("%s\n", "Failed to create lxcfs cgroup mountpoint.");
642 if (!umount_if_mounted()) {
643 lxcfs_error("%s\n", "Failed to clean up old lxcfs cgroup mountpoint.");
647 if (unshare(CLONE_NEWNS
) < 0) {
648 lxcfs_error("Failed to unshare mount namespace: %s.\n", strerror(errno
));
652 cgroup_ops
->mntns_fd
= preserve_ns(getpid(), "mnt");
653 if (cgroup_ops
->mntns_fd
< 0) {
654 lxcfs_error("Failed to preserve mount namespace: %s.\n", strerror(errno
));
658 if (mount(NULL
, "/", NULL
, MS_REC
| MS_PRIVATE
, 0) < 0) {
659 lxcfs_error("Failed to remount / private: %s.\n", strerror(errno
));
663 if (mount("tmpfs", BASEDIR
, "tmpfs", 0, "size=100000,mode=700") < 0) {
664 lxcfs_error("%s\n", "Failed to mount tmpfs over lxcfs cgroup mountpoint.");
671 static bool cgfs_mount_hierarchies(void)
673 if (!mkdir_p(BASEDIR DEFAULT_CGROUP_MOUNTPOINT
, 0755))
676 if (!cgroup_ops
->mount(cgroup_ops
, BASEDIR
))
679 for (struct hierarchy
**h
= cgroup_ops
->hierarchies
; h
&& *h
; h
++) {
680 __do_free
char *path
= must_make_path(BASEDIR
, (*h
)->mountpoint
, NULL
);
681 (*h
)->fd
= open(path
, O_DIRECTORY
| O_CLOEXEC
| O_NOFOLLOW
);
689 static bool cgfs_setup_controllers(void)
691 if (!cgfs_prepare_mounts())
694 if (!cgfs_mount_hierarchies())
695 return log_error_errno(false, errno
, "Failed to set up private lxcfs cgroup mounts");
703 static void __attribute__((constructor
)) lxcfs_init(void)
705 __do_close_prot_errno
int init_ns
= -EBADF
, pidfd
= -EBADF
;
709 char cwd
[MAXPATHLEN
];
711 cgroup_ops
= cgroup_init();
713 log_exit("Failed to initialize cgroup support");
715 /* Preserve initial namespace. */
717 init_ns
= preserve_ns(pid
, "mnt");
719 log_exit("Failed to preserve initial mount namespace");
721 cret
= getcwd(cwd
, MAXPATHLEN
);
723 log_exit("%s - Could not retrieve current working directory", strerror(errno
));
725 /* This function calls unshare(CLONE_NEWNS) our initial mount namespace
726 * to privately mount lxcfs cgroups. */
727 if (!cgfs_setup_controllers())
728 log_exit("Failed to setup private cgroup mounts for lxcfs");
730 if (setns(init_ns
, 0) < 0)
731 log_exit("%s - Failed to switch back to initial mount namespace", strerror(errno
));
733 if (!cret
|| chdir(cwd
) < 0)
734 log_exit("%s - Could not change back to original working directory", strerror(errno
));
737 log_exit("Failed to init CPU view");
739 fprintf(stderr
, "mount namespace: %d\n", cgroup_ops
->mntns_fd
);
740 fprintf(stderr
, "hierarchies:\n");
742 for (struct hierarchy
**h
= cgroup_ops
->hierarchies
; h
&& *h
; h
++, i
++) {
743 __do_free
char *controllers
= lxc_string_join(",", (const char **)(*h
)->controllers
, false);
744 fprintf(stderr
, " %2d: fd: %3d: %s\n", i
, (*h
)->fd
, controllers
?: "");
747 pidfd
= pidfd_open(pid
, 0);
748 if (pidfd
>= 0 && pidfd_send_signal(pidfd
, 0, NULL
, 0) == 0) {
749 can_use_pidfd
= true;
750 fprintf(stderr
, "Kernel supports pidfds\n");
753 fprintf(stderr
, "api_extensions:\n");
754 for (i
= 0; i
< nr_api_extensions
; i
++)
755 fprintf(stderr
, "- %s\n", api_extensions
[i
]);
758 static void __attribute__((destructor
)) lxcfs_exit(void)
760 lxcfs_debug("%s\n", "Running destructor for liblxcfs");
762 cgroup_exit(cgroup_ops
);