1 /* SPDX-License-Identifier: LGPL-2.1+ */
10 #include <linux/magic.h>
11 #include <linux/sched.h>
20 #include <sys/epoll.h>
22 #include <sys/mount.h>
23 #include <sys/param.h>
24 #include <sys/socket.h>
25 #include <sys/syscall.h>
26 #include <sys/sysinfo.h>
34 #include "api_extensions.h"
35 #include "cgroup_fuse.h"
36 #include "cgroups/cgroup.h"
37 #include "cgroups/cgroup_utils.h"
38 #include "memory_utils.h"
39 #include "proc_cpuview.h"
40 #include "syscall_numbers.h"
43 static bool can_use_pidfd
;
44 static bool can_use_swap
;
45 static bool can_use_sys_cpu
;
46 static bool has_versioned_opts
;
48 static volatile sig_atomic_t reload_successful
;
50 bool liblxcfs_functional(void)
52 return reload_successful
!= 0;
55 bool liblxcfs_can_use_swap(void)
60 bool liblxcfs_can_use_sys_cpu(void)
62 return can_use_sys_cpu
;
65 bool liblxcfs_has_versioned_opts(void)
67 return has_versioned_opts
;
70 /* Define pivot_root() if missing from the C library */
71 #ifndef HAVE_PIVOT_ROOT
72 static int pivot_root(const char *new_root
, const char *put_old
)
74 return syscall(__NR_pivot_root
, new_root
, put_old
);
77 extern int pivot_root(const char *new_root
, const char *put_old
);
81 * A table caching which pid is init for a pid namespace.
82 * When looking up which pid is init for $qpid, we first
83 * 1. Stat /proc/$qpid/ns/pid.
84 * 2. Check whether the ino_t is in our store.
85 * a. if not, fork a child in qpid's ns to send us
86 * ucred.pid = 1, and read the initpid. Cache
87 * initpid and creation time for /proc/initpid
88 * in a new store entry.
89 * b. if so, verify that /proc/initpid still matches
90 * what we have saved. If not, clear the store
91 * entry and go back to a. If so, return the
94 struct pidns_init_store
{
95 ino_t ino
; /* inode number for /proc/$pid/ns/pid */
96 pid_t initpid
; /* the pid of nit in that ns */
98 int64_t ctime
; /* the time at which /proc/$initpid was created */
99 struct pidns_init_store
*next
;
103 /* lol - look at how they are allocated in the kernel */
104 #define PIDNS_HASH_SIZE 4096
105 #define HASH(x) ((x) % PIDNS_HASH_SIZE)
107 static struct pidns_init_store
*pidns_hash_table
[PIDNS_HASH_SIZE
];
108 static pthread_mutex_t pidns_store_mutex
= PTHREAD_MUTEX_INITIALIZER
;
110 static void mutex_lock(pthread_mutex_t
*l
)
114 ret
= pthread_mutex_lock(l
);
116 log_exit("%s - returned %d\n", strerror(ret
), ret
);
119 struct cgroup_ops
*cgroup_ops
;
121 static void mutex_unlock(pthread_mutex_t
*l
)
125 ret
= pthread_mutex_unlock(l
);
127 log_exit("%s - returned %d\n", strerror(ret
), ret
);
130 static inline void store_lock(void)
132 mutex_lock(&pidns_store_mutex
);
135 static inline void store_unlock(void)
137 mutex_unlock(&pidns_store_mutex
);
142 * <pid-as-str> = INTTYPE_TO_STRLEN(pid_t)
146 #define LXCFS_PROC_PID_LEN \
147 (STRLITERALLEN("/proc/") + INTTYPE_TO_STRLEN(uint64_t) + +1)
149 static int initpid_still_valid_pidfd(struct pidns_init_store
*entry
)
153 if (entry
->init_pidfd
< 0)
154 return ret_errno(ENOSYS
);
156 ret
= pidfd_send_signal(entry
->init_pidfd
, 0, NULL
, 0);
159 return ret_errno(ENOSYS
);
167 static int initpid_still_valid_stat(struct pidns_init_store
*entry
)
170 char path
[LXCFS_PROC_PID_LEN
];
172 snprintf(path
, sizeof(path
), "/proc/%d", entry
->initpid
);
173 if (stat(path
, &st
) || entry
->ctime
!= st
.st_ctime
)
179 /* Must be called under store_lock */
180 static bool initpid_still_valid(struct pidns_init_store
*entry
)
184 ret
= initpid_still_valid_pidfd(entry
);
186 ret
= initpid_still_valid_stat(entry
);
191 /* Must be called under store_lock */
192 static void remove_initpid(struct pidns_init_store
*entry
)
194 struct pidns_init_store
*it
;
197 lxcfs_debug("Removing cached entry for pid %d from init pid cache",
200 ino_hash
= HASH(entry
->ino
);
201 if (pidns_hash_table
[ino_hash
] == entry
) {
202 pidns_hash_table
[ino_hash
] = entry
->next
;
203 close_prot_errno_disarm(entry
->init_pidfd
);
208 it
= pidns_hash_table
[ino_hash
];
210 if (it
->next
== entry
) {
211 it
->next
= entry
->next
;
212 close_prot_errno_disarm(entry
->init_pidfd
);
221 /* Must be called under store_lock */
222 static void prune_initpid_store(void)
224 static int64_t last_prune
= 0;
225 int64_t now
, threshold
;
228 last_prune
= time(NULL
);
233 if (now
< (last_prune
+ PURGE_SECS
))
236 lxcfs_debug("Pruning init pid cache");
239 threshold
= now
- 2 * PURGE_SECS
;
241 for (int i
= 0; i
< PIDNS_HASH_SIZE
; i
++) {
242 for (struct pidns_init_store
*entry
= pidns_hash_table
[i
], *prev
= NULL
; entry
;) {
243 if (entry
->lastcheck
< threshold
) {
244 struct pidns_init_store
*cur
= entry
;
246 lxcfs_debug("Removed cache entry for pid %d to init pid cache", cur
->initpid
);
249 prev
->next
= entry
->next
;
251 pidns_hash_table
[i
] = entry
->next
;
253 close_prot_errno_disarm(cur
->init_pidfd
);
263 static void clear_initpid_store(void)
266 for (int i
= 0; i
< PIDNS_HASH_SIZE
; i
++) {
267 for (struct pidns_init_store
*entry
= pidns_hash_table
[i
]; entry
;) {
268 struct pidns_init_store
*cur
= entry
;
270 lxcfs_debug("Removed cache entry for pid %d to init pid cache", cur
->initpid
);
272 pidns_hash_table
[i
] = entry
->next
;
274 close_prot_errno_disarm(cur
->init_pidfd
);
281 /* Must be called under store_lock */
282 static void save_initpid(ino_t pidns_inode
, pid_t pid
)
284 __do_free
struct pidns_init_store
*entry
= NULL
;
285 __do_close
int pidfd
= -EBADF
;
286 const struct lxcfs_opts
*opts
= fuse_get_context()->private_data
;
287 char path
[LXCFS_PROC_PID_LEN
];
291 if (opts
&& opts
->use_pidfd
&& can_use_pidfd
) {
292 pidfd
= pidfd_open(pid
, 0);
297 snprintf(path
, sizeof(path
), "/proc/%d", pid
);
301 entry
= zalloc(sizeof(*entry
));
305 ino_hash
= HASH(pidns_inode
);
306 *entry
= (struct pidns_init_store
){
309 .ctime
= st
.st_ctime
,
310 .next
= pidns_hash_table
[ino_hash
],
311 .lastcheck
= time(NULL
),
312 .init_pidfd
= move_fd(pidfd
),
314 pidns_hash_table
[ino_hash
] = move_ptr(entry
);
316 lxcfs_debug("Added cache entry %d for pid %d to init pid cache", ino_hash
, pid
);
320 * Given the stat(2) info for a nsfd pid inode, lookup the init_pid_store
321 * entry for the inode number and creation time. Verify that the init pid
322 * is still valid. If not, remove it. Return the entry if valid, NULL
324 * Must be called under store_lock
326 static pid_t
lookup_verify_initpid(ino_t pidns_inode
)
328 struct pidns_init_store
*entry
= pidns_hash_table
[HASH(pidns_inode
)];
331 if (entry
->ino
== pidns_inode
) {
332 if (initpid_still_valid(entry
)) {
333 entry
->lastcheck
= time(NULL
);
334 return entry
->initpid
;
337 remove_initpid(entry
);
338 return ret_errno(ESRCH
);
343 return ret_errno(ESRCH
);
346 static bool send_creds_ok(int sock_fd
)
348 char v
= '1'; /* we are the child */
349 struct ucred cred
= {
355 return send_creds(sock_fd
, &cred
, v
, true) == SEND_CREDS_OK
;
358 __returns_twice pid_t
lxcfs_raw_clone(unsigned long flags
, int *pidfd
)
361 * These flags don't interest at all so we don't jump through any hoops
362 * of retrieving them and passing them to the kernel.
365 if ((flags
& (CLONE_VM
| CLONE_PARENT_SETTID
| CLONE_CHILD_SETTID
|
366 CLONE_CHILD_CLEARTID
| CLONE_SETTLS
)))
369 #if defined(__s390x__) || defined(__s390__) || defined(__CRIS__)
370 /* On s390/s390x and cris the order of the first and second arguments
371 * of the system call is reversed.
373 return syscall(__NR_clone
, NULL
, flags
| SIGCHLD
, pidfd
);
374 #elif defined(__sparc__) && defined(__arch64__)
377 * sparc64 always returns the other process id in %o0, and a
378 * boolean flag whether this is the child or the parent in %o1.
379 * Inline assembly is needed to get the flag returned in %o1.
381 register long g1
asm("g1") = __NR_clone
;
382 register long o0
asm("o0") = flags
| SIGCHLD
;
383 register long o1
asm("o1") = 0; /* is parent/child indicator */
384 register long o2
asm("o2") = (unsigned long)pidfd
;
385 long is_error
, retval
, in_child
;
389 #if defined(__arch64__)
390 "t 0x6d\n\t" /* 64-bit trap */
392 "t 0x10\n\t" /* 32-bit trap */
395 * catch errors: On sparc, the carry bit (csr) in the
396 * processor status register (psr) is used instead of a
400 : "=r"(g1
), "=r"(o0
), "=r"(o1
), "=r"(o2
) /* outputs */
401 : "r"(g1
), "r"(o0
), "r"(o1
), "r"(o2
) /* inputs */
402 : "%cc"); /* clobbers */
419 #elif defined(__ia64__)
420 /* On ia64 the stack and stack size are passed as separate arguments. */
421 return syscall(__NR_clone
, flags
| SIGCHLD
, NULL
, prctl_arg(0), pidfd
);
423 return syscall(__NR_clone
, flags
| SIGCHLD
, NULL
, pidfd
);
427 #define LXCFS_PROC_PID_NS_LEN \
428 (STRLITERALLEN("/proc/") + INTTYPE_TO_STRLEN(uint64_t) + \
429 STRLITERALLEN("/ns/pid") + 1)
432 * clone a task which switches to @task's namespace and writes '1'.
433 * over a unix sock so we can read the task's reaper's pid in our
436 * Note: glibc's fork() does not respect pidns, which can lead to failed
437 * assertions inside glibc (and thus failed forks) if the child's pid in
438 * the pidns and the parent pid outside are identical. Using clone prevents
441 static void write_task_init_pid_exit(int sock
, pid_t target
)
443 __do_close
int fd
= -EBADF
;
444 char path
[LXCFS_PROC_PID_NS_LEN
];
447 snprintf(path
, sizeof(path
), "/proc/%d/ns/pid", (int)target
);
448 fd
= open(path
, O_RDONLY
| O_CLOEXEC
);
450 log_exit("write_task_init_pid_exit open of ns/pid");
453 log_exit("Failed to setns to pid namespace of process %d", target
);
455 pid
= lxcfs_raw_clone(0, NULL
);
460 if (!send_creds_ok(sock
))
466 if (!wait_for_pid(pid
))
472 static pid_t
scm_init_pid(pid_t task
)
476 struct ucred cred
= {
484 if (socketpair(AF_UNIX
, SOCK_DGRAM
, 0, sock
) < 0)
493 write_task_init_pid_exit(sock
[0], task
);
497 if (!recv_creds(sock
[1], &cred
, &v
))
511 pid_t
lookup_initpid_in_store(pid_t pid
)
513 pid_t hashed_pid
= 0;
514 char path
[LXCFS_PROC_PID_NS_LEN
];
517 snprintf(path
, sizeof(path
), "/proc/%d/ns/pid", pid
);
519 return ret_errno(ESRCH
);
523 hashed_pid
= lookup_verify_initpid(st
.st_ino
);
524 if (hashed_pid
< 0) {
525 /* release the mutex as the following call is expensive */
528 hashed_pid
= scm_init_pid(pid
);
533 save_initpid(st
.st_ino
, hashed_pid
);
537 * Prune at the end in case we're pruning the value
538 * we were about to return.
540 prune_initpid_store();
547 * Functions needed to setup cgroups in the __constructor__.
550 static bool umount_if_mounted(void)
552 if (umount2(BASEDIR
, MNT_DETACH
) < 0 && errno
!= EINVAL
) {
553 lxcfs_error("Failed to unmount %s: %s.\n", BASEDIR
, strerror(errno
));
559 /* __typeof__ should be safe to use with all compilers. */
560 typedef __typeof__(((struct statfs
*)NULL
)->f_type
) fs_type_magic
;
561 static bool has_fs_type(const struct statfs
*fs
, fs_type_magic magic_val
)
563 return (fs
->f_type
== (fs_type_magic
)magic_val
);
567 * looking at fs/proc_namespace.c, it appears we can
568 * actually expect the rootfs entry to very specifically contain
569 * " - rootfs rootfs "
570 * IIUC, so long as we've chrooted so that rootfs is not our root,
571 * the rootfs entry should always be skipped in mountinfo contents.
573 static bool is_on_ramfs(void)
575 __do_free
char *line
= NULL
;
576 __do_free
void *fopen_cache
= NULL
;
577 __do_fclose
FILE *f
= NULL
;
580 f
= fopen_cached("/proc/self/mountinfo", "re", &fopen_cache
);
584 while (getline(&line
, &len
, f
) != -1) {
588 for (p
= line
, i
= 0; p
&& i
< 4; i
++)
589 p
= strchr(p
+ 1, ' ');
593 p2
= strchr(p
+ 1, ' ');
597 if (strcmp(p
+ 1, "/") == 0) {
598 /* This is '/'. Is it the ramfs? */
599 p
= strchr(p2
+ 1, '-');
600 if (p
&& strncmp(p
, "- rootfs rootfs ", 16) == 0)
608 static int pivot_enter(void)
610 __do_close
int oldroot
= -EBADF
, newroot
= -EBADF
;
612 oldroot
= open("/", O_DIRECTORY
| O_RDONLY
| O_CLOEXEC
);
614 return log_error_errno(-1, errno
,
615 "Failed to open old root for fchdir");
617 newroot
= open(ROOTDIR
, O_DIRECTORY
| O_RDONLY
| O_CLOEXEC
);
619 return log_error_errno(-1, errno
,
620 "Failed to open new root for fchdir");
622 /* change into new root fs */
623 if (fchdir(newroot
) < 0)
624 return log_error_errno(-1,
625 errno
, "Failed to change directory to new rootfs: %s",
628 /* pivot_root into our new root fs */
629 if (pivot_root(".", ".") < 0)
630 return log_error_errno(-1, errno
,
631 "pivot_root() syscall failed: %s",
635 * At this point the old-root is mounted on top of our new-root.
636 * To unmounted it we must not be chdir'd into it, so escape back
639 if (fchdir(oldroot
) < 0)
640 return log_error_errno(-1, errno
, "Failed to enter old root");
642 if (umount2(".", MNT_DETACH
) < 0)
643 return log_error_errno(-1, errno
, "Failed to detach old root");
645 if (fchdir(newroot
) < 0)
646 return log_error_errno(-1, errno
, "Failed to re-enter new root");
651 static int chroot_enter(void)
653 if (mount(ROOTDIR
, "/", NULL
, MS_REC
| MS_BIND
, NULL
)) {
654 lxcfs_error("Failed to recursively bind-mount %s into /.", ROOTDIR
);
658 if (chroot(".") < 0) {
659 lxcfs_error("Call to chroot() failed: %s.\n", strerror(errno
));
663 if (chdir("/") < 0) {
664 lxcfs_error("Failed to change directory: %s.\n", strerror(errno
));
671 static int permute_and_enter(void)
675 if (statfs("/", &sb
) < 0) {
676 lxcfs_error("%s\n", "Could not stat / mountpoint.");
680 /* has_fs_type() is not reliable. When the ramfs is a tmpfs it will
681 * likely report TMPFS_MAGIC. Hence, when it reports no we still check
682 * /proc/1/mountinfo. */
683 if (has_fs_type(&sb
, RAMFS_MAGIC
) || is_on_ramfs())
684 return chroot_enter();
686 if (pivot_enter() < 0) {
687 lxcfs_error("%s\n", "Could not perform pivot root.");
694 /* Prepare our new clean root. */
695 static int permute_prepare(void)
697 if (mkdir(ROOTDIR
, 0700) < 0 && errno
!= EEXIST
) {
698 lxcfs_error("%s\n", "Failed to create directory for new root.");
702 if (mount("/", ROOTDIR
, NULL
, MS_BIND
, 0) < 0) {
703 lxcfs_error("Failed to bind-mount / for new root: %s.\n", strerror(errno
));
707 if (mount(RUNTIME_PATH
, ROOTDIR RUNTIME_PATH
, NULL
, MS_BIND
, 0) < 0) {
708 lxcfs_error("Failed to bind-mount /run into new root: %s.\n", strerror(errno
));
712 if (mount(BASEDIR
, ROOTDIR BASEDIR
, NULL
, MS_REC
| MS_MOVE
, 0) < 0) {
713 printf("Failed to move " BASEDIR
" into new root: %s.\n", strerror(errno
));
720 /* Calls chroot() on ramfs, pivot_root() in all other cases. */
721 static bool permute_root(void)
723 /* Prepare new root. */
724 if (permute_prepare() < 0)
727 /* Pivot into new root. */
728 if (permute_and_enter() < 0)
734 static bool cgfs_prepare_mounts(void)
736 if (!mkdir_p(BASEDIR
, 0700)) {
737 lxcfs_error("%s\n", "Failed to create lxcfs cgroup mountpoint.");
741 if (!umount_if_mounted()) {
742 lxcfs_error("%s\n", "Failed to clean up old lxcfs cgroup mountpoint.");
746 if (unshare(CLONE_NEWNS
) < 0) {
747 lxcfs_error("Failed to unshare mount namespace: %s.\n", strerror(errno
));
751 cgroup_ops
->mntns_fd
= preserve_ns(getpid(), "mnt");
752 if (cgroup_ops
->mntns_fd
< 0) {
753 lxcfs_error("Failed to preserve mount namespace: %s.\n", strerror(errno
));
757 if (mount(NULL
, "/", NULL
, MS_REC
| MS_PRIVATE
, 0) < 0) {
758 lxcfs_error("Failed to remount / private: %s.\n", strerror(errno
));
762 if (mount("tmpfs", BASEDIR
, "tmpfs", 0, "size=100000,mode=700") < 0) {
763 lxcfs_error("%s\n", "Failed to mount tmpfs over lxcfs cgroup mountpoint.");
770 static bool cgfs_mount_hierarchies(void)
772 if (!mkdir_p(BASEDIR DEFAULT_CGROUP_MOUNTPOINT
, 0755))
775 if (!cgroup_ops
->mount(cgroup_ops
, BASEDIR
))
778 for (struct hierarchy
**h
= cgroup_ops
->hierarchies
; h
&& *h
; h
++) {
779 __do_free
char *path
= must_make_path(BASEDIR
, (*h
)->mountpoint
, NULL
);
780 (*h
)->fd
= open(path
, O_DIRECTORY
| O_CLOEXEC
| O_NOFOLLOW
);
788 static bool cgfs_setup_controllers(void)
790 if (!cgfs_prepare_mounts())
793 if (!cgfs_mount_hierarchies())
794 return log_error_errno(false, errno
, "Failed to set up private lxcfs cgroup mounts");
802 static void sigusr2_toggle_virtualization(int signo
, siginfo_t
*info
, void *extra
)
806 if (reload_successful
) {
807 reload_successful
= 0;
809 /* write() is async signal safe */
810 ret
= write(STDERR_FILENO
,
811 "Switched into non-virtualization mode\n",
812 STRLITERALLEN("Switched into non-virtualization mode\n"));
814 goto please_compiler
;
816 reload_successful
= 1;
818 /* write() is async signal safe */
819 ret
= write(STDERR_FILENO
, "Switched into virtualization mode\n",
820 STRLITERALLEN("Switched into virtualization mode\n"));
822 goto please_compiler
;
827 * The write() syscall is a function whose return value needs to be
828 * checked. Otherwise the compiler will warn.Another one could be to
829 * use syscall(__NR_write, ...) directly but whatever.
834 static void __attribute__((constructor
)) lxcfs_init(void)
836 __do_close
int init_ns
= -EBADF
, root_fd
= -EBADF
,
841 lxcfs_info("Running constructor %s to reload liblxcfs", __func__
);
843 cgroup_ops
= cgroup_init();
845 lxcfs_info("Failed to initialize cgroup support");
849 /* Preserve initial namespace. */
851 init_ns
= preserve_ns(pid
, "mnt");
853 lxcfs_info("Failed to preserve initial mount namespace");
857 /* This function calls unshare(CLONE_NEWNS) our initial mount namespace
858 * to privately mount lxcfs cgroups. */
859 if (!cgfs_setup_controllers()) {
860 log_exit("Failed to setup private cgroup mounts for lxcfs");
864 if (setns(init_ns
, 0) < 0) {
865 log_exit("%s - Failed to switch back to initial mount namespace", strerror(errno
));
869 if (!init_cpuview()) {
870 log_exit("Failed to init CPU view");
874 lxcfs_info("mount namespace: %d", cgroup_ops
->mntns_fd
);
875 lxcfs_info("hierarchies:");
877 for (struct hierarchy
**h
= cgroup_ops
->hierarchies
; h
&& *h
; h
++, i
++) {
878 char **controller_list
= (*h
)->controllers
;
879 __do_free
char *controllers
= NULL
;
880 if (controller_list
&& *controller_list
)
881 controllers
= lxc_string_join(",", (const char **)controller_list
, false);
882 lxcfs_info(" %2d: fd: %3d: %s", i
, (*h
)->fd
, controllers
?: "");
885 pidfd
= pidfd_open(pid
, 0);
886 if (pidfd
>= 0 && pidfd_send_signal(pidfd
, 0, NULL
, 0) == 0) {
887 can_use_pidfd
= true;
888 lxcfs_info("Kernel supports pidfds");
891 can_use_swap
= cgroup_ops
->can_use_swap(cgroup_ops
);
893 lxcfs_info("Kernel supports swap accounting");
895 lxcfs_info("Kernel does not support swap accounting");
897 lxcfs_info("api_extensions:");
898 for (size_t nr
= 0; nr
< nr_api_extensions
; nr
++)
899 lxcfs_info("- %s", api_extensions
[nr
]);
901 root_fd
= open("/", O_PATH
| O_CLOEXEC
);
903 lxcfs_info("%s - Failed to open root directory", strerror(errno
));
904 else if (fchdir(root_fd
) < 0)
905 lxcfs_info("%s - Failed to change to root directory", strerror(errno
));
907 if (install_signal_handler(SIGUSR2
, sigusr2_toggle_virtualization
)) {
908 lxcfs_info("%s - Failed to install SIGUSR2 signal handler", strerror(errno
));
912 reload_successful
= 1;
916 reload_successful
= 0;
917 lxcfs_info("Failed to run constructor %s to reload liblxcfs", __func__
);
920 static void __attribute__((destructor
)) lxcfs_exit(void)
922 lxcfs_info("Running destructor %s", __func__
);
924 clear_initpid_store();
926 cgroup_exit(cgroup_ops
);
929 void *lxcfs_fuse_init(struct fuse_conn_info
*conn
, void *data
)
931 struct fuse_context
*fc
= fuse_get_context();
932 can_use_sys_cpu
= true;
933 has_versioned_opts
= true;
934 return fc
->private_data
;