2 * POSIX message queues filesystem for Linux.
4 * Copyright (C) 2003,2004 Krzysztof Benedyczak (golbi@mat.uni.torun.pl)
5 * Michal Wronski (michal.wronski@gmail.com)
7 * Spinlocks: Mohamed Abbas (abbas.mohamed@intel.com)
8 * Lockless receive & send, fd based notify:
9 * Manfred Spraul (manfred@colorfullife.com)
11 * Audit: George Wilson (ltcgcw@us.ibm.com)
13 * This file is released under the GPL.
16 #include <linux/capability.h>
17 #include <linux/init.h>
18 #include <linux/pagemap.h>
19 #include <linux/file.h>
20 #include <linux/mount.h>
21 #include <linux/fs_context.h>
22 #include <linux/namei.h>
23 #include <linux/sysctl.h>
24 #include <linux/poll.h>
25 #include <linux/mqueue.h>
26 #include <linux/msg.h>
27 #include <linux/skbuff.h>
28 #include <linux/vmalloc.h>
29 #include <linux/netlink.h>
30 #include <linux/syscalls.h>
31 #include <linux/audit.h>
32 #include <linux/signal.h>
33 #include <linux/mutex.h>
34 #include <linux/nsproxy.h>
35 #include <linux/pid.h>
36 #include <linux/ipc_namespace.h>
37 #include <linux/user_namespace.h>
38 #include <linux/slab.h>
39 #include <linux/sched/wake_q.h>
40 #include <linux/sched/signal.h>
41 #include <linux/sched/user.h>
46 struct mqueue_fs_context {
47 struct ipc_namespace *ipc_ns;
48 bool newns; /* Set if newly created ipc namespace */
51 #define MQUEUE_MAGIC 0x19800202
52 #define DIRENT_SIZE 20
53 #define FILENT_SIZE 80
61 struct posix_msg_tree_node {
62 struct rb_node rb_node;
63 struct list_head msg_list;
70 * Accesses to a message queue are synchronized by acquiring info->lock.
72 * There are two notable exceptions:
73 * - The actual wakeup of a sleeping task is performed using the wake_q
74 * framework. info->lock is already released when wake_up_q is called.
75 * - The exit codepaths after sleeping check ext_wait_queue->state without
76 * any locks. If it is STATE_READY, then the syscall is completed without
77 * acquiring info->lock.
80 * To achieve proper release/acquire memory barrier pairing, the state is set to
81 * STATE_READY with smp_store_release(), and it is read with READ_ONCE followed
82 * by smp_acquire__after_ctrl_dep(). In addition, wake_q_add_safe() is used.
84 * This prevents the following races:
86 * 1) With the simple wake_q_add(), the task could be gone already before
87 * the increase of the reference happens
90 * WRITE_ONCE(wait.state, STATE_NONE);
91 * schedule_hrtimeout()
93 * if (cmpxchg()) // success
94 * ->state = STATE_READY (reordered)
96 * if (wait.state == STATE_READY) return;
97 * sysret to user space
99 * get_task_struct() // UaF
101 * Solution: Use wake_q_add_safe() and perform the get_task_struct() before
102 * the smp_store_release() that does ->state = STATE_READY.
104 * 2) Without proper _release/_acquire barriers, the woken up task
105 * could read stale data
110 * WRITE_ONCE(wait.state, STATE_NONE);
111 * schedule_hrtimeout()
112 * state = STATE_READY;
114 * if (wait.state == STATE_READY) return;
115 * msg_ptr = wait.msg; // Access to stale data!
116 * receiver->msg = message; (reordered)
118 * Solution: use _release and _acquire barriers.
120 * 3) There is intentionally no barrier when setting current->state
121 * to TASK_INTERRUPTIBLE: spin_unlock(&info->lock) provides the
122 * release memory barrier, and the wakeup is triggered when holding
123 * info->lock, i.e. spin_lock(&info->lock) provided a pairing
124 * acquire memory barrier.
127 struct ext_wait_queue { /* queue of sleeping tasks */
128 struct task_struct *task;
129 struct list_head list;
130 struct msg_msg *msg; /* ptr of loaded message */
131 int state; /* one of STATE_* values */
134 struct mqueue_inode_info {
136 struct inode vfs_inode;
137 wait_queue_head_t wait_q;
139 struct rb_root msg_tree;
140 struct rb_node *msg_tree_rightmost;
141 struct posix_msg_tree_node *node_cache;
144 struct sigevent notify;
145 struct pid *notify_owner;
146 u32 notify_self_exec_id;
147 struct user_namespace *notify_user_ns;
148 struct ucounts *ucounts; /* user who created, for accounting */
149 struct sock *notify_sock;
150 struct sk_buff *notify_cookie;
152 /* for tasks waiting for free space and messages, respectively */
153 struct ext_wait_queue e_wait_q[2];
155 unsigned long qsize; /* size of queue in memory (sum of all msgs) */
158 static struct file_system_type mqueue_fs_type;
159 static const struct inode_operations mqueue_dir_inode_operations;
160 static const struct file_operations mqueue_file_operations;
161 static const struct super_operations mqueue_super_ops;
162 static const struct fs_context_operations mqueue_fs_context_ops;
163 static void remove_notification(struct mqueue_inode_info *info);
165 static struct kmem_cache *mqueue_inode_cachep;
167 static inline struct mqueue_inode_info *MQUEUE_I(struct inode *inode)
169 return container_of(inode, struct mqueue_inode_info, vfs_inode);
173 * This routine should be called with the mq_lock held.
175 static inline struct ipc_namespace *__get_ns_from_inode(struct inode *inode)
177 return get_ipc_ns(inode->i_sb->s_fs_info);
180 static struct ipc_namespace *get_ns_from_inode(struct inode *inode)
182 struct ipc_namespace *ns;
185 ns = __get_ns_from_inode(inode);
186 spin_unlock(&mq_lock);
190 /* Auxiliary functions to manipulate messages' list */
191 static int msg_insert(struct msg_msg *msg, struct mqueue_inode_info *info)
193 struct rb_node **p, *parent = NULL;
194 struct posix_msg_tree_node *leaf;
195 bool rightmost = true;
197 p = &info->msg_tree.rb_node;
200 leaf = rb_entry(parent, struct posix_msg_tree_node, rb_node);
202 if (likely(leaf->priority == msg->m_type))
204 else if (msg->m_type < leaf->priority) {
210 if (info->node_cache) {
211 leaf = info->node_cache;
212 info->node_cache = NULL;
214 leaf = kmalloc(sizeof(*leaf), GFP_ATOMIC);
217 INIT_LIST_HEAD(&leaf->msg_list);
219 leaf->priority = msg->m_type;
222 info->msg_tree_rightmost = &leaf->rb_node;
224 rb_link_node(&leaf->rb_node, parent, p);
225 rb_insert_color(&leaf->rb_node, &info->msg_tree);
227 info->attr.mq_curmsgs++;
228 info->qsize += msg->m_ts;
229 list_add_tail(&msg->m_list, &leaf->msg_list);
233 static inline void msg_tree_erase(struct posix_msg_tree_node *leaf,
234 struct mqueue_inode_info *info)
236 struct rb_node *node = &leaf->rb_node;
238 if (info->msg_tree_rightmost == node)
239 info->msg_tree_rightmost = rb_prev(node);
241 rb_erase(node, &info->msg_tree);
242 if (info->node_cache)
245 info->node_cache = leaf;
248 static inline struct msg_msg *msg_get(struct mqueue_inode_info *info)
250 struct rb_node *parent = NULL;
251 struct posix_msg_tree_node *leaf;
256 * During insert, low priorities go to the left and high to the
257 * right. On receive, we want the highest priorities first, so
258 * walk all the way to the right.
260 parent = info->msg_tree_rightmost;
262 if (info->attr.mq_curmsgs) {
263 pr_warn_once("Inconsistency in POSIX message queue, "
264 "no tree element, but supposedly messages "
266 info->attr.mq_curmsgs = 0;
270 leaf = rb_entry(parent, struct posix_msg_tree_node, rb_node);
271 if (unlikely(list_empty(&leaf->msg_list))) {
272 pr_warn_once("Inconsistency in POSIX message queue, "
273 "empty leaf node but we haven't implemented "
274 "lazy leaf delete!\n");
275 msg_tree_erase(leaf, info);
278 msg = list_first_entry(&leaf->msg_list,
279 struct msg_msg, m_list);
280 list_del(&msg->m_list);
281 if (list_empty(&leaf->msg_list)) {
282 msg_tree_erase(leaf, info);
285 info->attr.mq_curmsgs--;
286 info->qsize -= msg->m_ts;
290 static struct inode *mqueue_get_inode(struct super_block *sb,
291 struct ipc_namespace *ipc_ns, umode_t mode,
292 struct mq_attr *attr)
297 inode = new_inode(sb);
301 inode->i_ino = get_next_ino();
302 inode->i_mode = mode;
303 inode->i_uid = current_fsuid();
304 inode->i_gid = current_fsgid();
305 inode->i_mtime = inode->i_ctime = inode->i_atime = current_time(inode);
308 struct mqueue_inode_info *info;
309 unsigned long mq_bytes, mq_treesize;
311 inode->i_fop = &mqueue_file_operations;
312 inode->i_size = FILENT_SIZE;
313 /* mqueue specific info */
314 info = MQUEUE_I(inode);
315 spin_lock_init(&info->lock);
316 init_waitqueue_head(&info->wait_q);
317 INIT_LIST_HEAD(&info->e_wait_q[0].list);
318 INIT_LIST_HEAD(&info->e_wait_q[1].list);
319 info->notify_owner = NULL;
320 info->notify_user_ns = NULL;
322 info->ucounts = NULL; /* set when all is ok */
323 info->msg_tree = RB_ROOT;
324 info->msg_tree_rightmost = NULL;
325 info->node_cache = NULL;
326 memset(&info->attr, 0, sizeof(info->attr));
327 info->attr.mq_maxmsg = min(ipc_ns->mq_msg_max,
328 ipc_ns->mq_msg_default);
329 info->attr.mq_msgsize = min(ipc_ns->mq_msgsize_max,
330 ipc_ns->mq_msgsize_default);
332 info->attr.mq_maxmsg = attr->mq_maxmsg;
333 info->attr.mq_msgsize = attr->mq_msgsize;
336 * We used to allocate a static array of pointers and account
337 * the size of that array as well as one msg_msg struct per
338 * possible message into the queue size. That's no longer
339 * accurate as the queue is now an rbtree and will grow and
340 * shrink depending on usage patterns. We can, however, still
341 * account one msg_msg struct per message, but the nodes are
342 * allocated depending on priority usage, and most programs
343 * only use one, or a handful, of priorities. However, since
344 * this is pinned memory, we need to assume worst case, so
345 * that means the min(mq_maxmsg, max_priorities) * struct
346 * posix_msg_tree_node.
350 if (info->attr.mq_maxmsg <= 0 || info->attr.mq_msgsize <= 0)
352 if (capable(CAP_SYS_RESOURCE)) {
353 if (info->attr.mq_maxmsg > HARD_MSGMAX ||
354 info->attr.mq_msgsize > HARD_MSGSIZEMAX)
357 if (info->attr.mq_maxmsg > ipc_ns->mq_msg_max ||
358 info->attr.mq_msgsize > ipc_ns->mq_msgsize_max)
362 /* check for overflow */
363 if (info->attr.mq_msgsize > ULONG_MAX/info->attr.mq_maxmsg)
365 mq_treesize = info->attr.mq_maxmsg * sizeof(struct msg_msg) +
366 min_t(unsigned int, info->attr.mq_maxmsg, MQ_PRIO_MAX) *
367 sizeof(struct posix_msg_tree_node);
368 mq_bytes = info->attr.mq_maxmsg * info->attr.mq_msgsize;
369 if (mq_bytes + mq_treesize < mq_bytes)
371 mq_bytes += mq_treesize;
372 info->ucounts = get_ucounts(current_ucounts());
377 msgqueue = inc_rlimit_ucounts(info->ucounts, UCOUNT_RLIMIT_MSGQUEUE, mq_bytes);
378 if (msgqueue == LONG_MAX || msgqueue > rlimit(RLIMIT_MSGQUEUE)) {
379 dec_rlimit_ucounts(info->ucounts, UCOUNT_RLIMIT_MSGQUEUE, mq_bytes);
380 spin_unlock(&mq_lock);
381 put_ucounts(info->ucounts);
382 info->ucounts = NULL;
383 /* mqueue_evict_inode() releases info->messages */
387 spin_unlock(&mq_lock);
389 } else if (S_ISDIR(mode)) {
391 /* Some things misbehave if size == 0 on a directory */
392 inode->i_size = 2 * DIRENT_SIZE;
393 inode->i_op = &mqueue_dir_inode_operations;
394 inode->i_fop = &simple_dir_operations;
404 static int mqueue_fill_super(struct super_block *sb, struct fs_context *fc)
407 struct ipc_namespace *ns = sb->s_fs_info;
409 sb->s_iflags |= SB_I_NOEXEC | SB_I_NODEV;
410 sb->s_blocksize = PAGE_SIZE;
411 sb->s_blocksize_bits = PAGE_SHIFT;
412 sb->s_magic = MQUEUE_MAGIC;
413 sb->s_op = &mqueue_super_ops;
415 inode = mqueue_get_inode(sb, ns, S_IFDIR | S_ISVTX | S_IRWXUGO, NULL);
417 return PTR_ERR(inode);
419 sb->s_root = d_make_root(inode);
425 static int mqueue_get_tree(struct fs_context *fc)
427 struct mqueue_fs_context *ctx = fc->fs_private;
430 * With a newly created ipc namespace, we don't need to do a search
431 * for an ipc namespace match, but we still need to set s_fs_info.
434 fc->s_fs_info = ctx->ipc_ns;
435 return get_tree_nodev(fc, mqueue_fill_super);
437 return get_tree_keyed(fc, mqueue_fill_super, ctx->ipc_ns);
440 static void mqueue_fs_context_free(struct fs_context *fc)
442 struct mqueue_fs_context *ctx = fc->fs_private;
444 put_ipc_ns(ctx->ipc_ns);
448 static int mqueue_init_fs_context(struct fs_context *fc)
450 struct mqueue_fs_context *ctx;
452 ctx = kzalloc(sizeof(struct mqueue_fs_context), GFP_KERNEL);
456 ctx->ipc_ns = get_ipc_ns(current->nsproxy->ipc_ns);
457 put_user_ns(fc->user_ns);
458 fc->user_ns = get_user_ns(ctx->ipc_ns->user_ns);
459 fc->fs_private = ctx;
460 fc->ops = &mqueue_fs_context_ops;
465 * mq_init_ns() is currently the only caller of mq_create_mount().
466 * So the ns parameter is always a newly created ipc namespace.
468 static struct vfsmount *mq_create_mount(struct ipc_namespace *ns)
470 struct mqueue_fs_context *ctx;
471 struct fs_context *fc;
472 struct vfsmount *mnt;
474 fc = fs_context_for_mount(&mqueue_fs_type, SB_KERNMOUNT);
478 ctx = fc->fs_private;
480 put_ipc_ns(ctx->ipc_ns);
481 ctx->ipc_ns = get_ipc_ns(ns);
482 put_user_ns(fc->user_ns);
483 fc->user_ns = get_user_ns(ctx->ipc_ns->user_ns);
490 static void init_once(void *foo)
492 struct mqueue_inode_info *p = foo;
494 inode_init_once(&p->vfs_inode);
497 static struct inode *mqueue_alloc_inode(struct super_block *sb)
499 struct mqueue_inode_info *ei;
501 ei = alloc_inode_sb(sb, mqueue_inode_cachep, GFP_KERNEL);
504 return &ei->vfs_inode;
507 static void mqueue_free_inode(struct inode *inode)
509 kmem_cache_free(mqueue_inode_cachep, MQUEUE_I(inode));
512 static void mqueue_evict_inode(struct inode *inode)
514 struct mqueue_inode_info *info;
515 struct ipc_namespace *ipc_ns;
516 struct msg_msg *msg, *nmsg;
521 if (S_ISDIR(inode->i_mode))
524 ipc_ns = get_ns_from_inode(inode);
525 info = MQUEUE_I(inode);
526 spin_lock(&info->lock);
527 while ((msg = msg_get(info)) != NULL)
528 list_add_tail(&msg->m_list, &tmp_msg);
529 kfree(info->node_cache);
530 spin_unlock(&info->lock);
532 list_for_each_entry_safe(msg, nmsg, &tmp_msg, m_list) {
533 list_del(&msg->m_list);
538 unsigned long mq_bytes, mq_treesize;
540 /* Total amount of bytes accounted for the mqueue */
541 mq_treesize = info->attr.mq_maxmsg * sizeof(struct msg_msg) +
542 min_t(unsigned int, info->attr.mq_maxmsg, MQ_PRIO_MAX) *
543 sizeof(struct posix_msg_tree_node);
545 mq_bytes = mq_treesize + (info->attr.mq_maxmsg *
546 info->attr.mq_msgsize);
549 dec_rlimit_ucounts(info->ucounts, UCOUNT_RLIMIT_MSGQUEUE, mq_bytes);
551 * get_ns_from_inode() ensures that the
552 * (ipc_ns = sb->s_fs_info) is either a valid ipc_ns
553 * to which we now hold a reference, or it is NULL.
554 * We can't put it here under mq_lock, though.
557 ipc_ns->mq_queues_count--;
558 spin_unlock(&mq_lock);
559 put_ucounts(info->ucounts);
560 info->ucounts = NULL;
566 static int mqueue_create_attr(struct dentry *dentry, umode_t mode, void *arg)
568 struct inode *dir = dentry->d_parent->d_inode;
570 struct mq_attr *attr = arg;
572 struct ipc_namespace *ipc_ns;
575 ipc_ns = __get_ns_from_inode(dir);
581 if (ipc_ns->mq_queues_count >= ipc_ns->mq_queues_max &&
582 !capable(CAP_SYS_RESOURCE)) {
586 ipc_ns->mq_queues_count++;
587 spin_unlock(&mq_lock);
589 inode = mqueue_get_inode(dir->i_sb, ipc_ns, mode, attr);
591 error = PTR_ERR(inode);
593 ipc_ns->mq_queues_count--;
598 dir->i_size += DIRENT_SIZE;
599 dir->i_ctime = dir->i_mtime = dir->i_atime = current_time(dir);
601 d_instantiate(dentry, inode);
605 spin_unlock(&mq_lock);
611 static int mqueue_create(struct user_namespace *mnt_userns, struct inode *dir,
612 struct dentry *dentry, umode_t mode, bool excl)
614 return mqueue_create_attr(dentry, mode, NULL);
617 static int mqueue_unlink(struct inode *dir, struct dentry *dentry)
619 struct inode *inode = d_inode(dentry);
621 dir->i_ctime = dir->i_mtime = dir->i_atime = current_time(dir);
622 dir->i_size -= DIRENT_SIZE;
629 * This is routine for system read from queue file.
630 * To avoid mess with doing here some sort of mq_receive we allow
631 * to read only queue size & notification info (the only values
632 * that are interesting from user point of view and aren't accessible
633 * through std routines)
635 static ssize_t mqueue_read_file(struct file *filp, char __user *u_data,
636 size_t count, loff_t *off)
638 struct mqueue_inode_info *info = MQUEUE_I(file_inode(filp));
639 char buffer[FILENT_SIZE];
642 spin_lock(&info->lock);
643 snprintf(buffer, sizeof(buffer),
644 "QSIZE:%-10lu NOTIFY:%-5d SIGNO:%-5d NOTIFY_PID:%-6d\n",
646 info->notify_owner ? info->notify.sigev_notify : 0,
647 (info->notify_owner &&
648 info->notify.sigev_notify == SIGEV_SIGNAL) ?
649 info->notify.sigev_signo : 0,
650 pid_vnr(info->notify_owner));
651 spin_unlock(&info->lock);
652 buffer[sizeof(buffer)-1] = '\0';
654 ret = simple_read_from_buffer(u_data, count, off, buffer,
659 file_inode(filp)->i_atime = file_inode(filp)->i_ctime = current_time(file_inode(filp));
663 static int mqueue_flush_file(struct file *filp, fl_owner_t id)
665 struct mqueue_inode_info *info = MQUEUE_I(file_inode(filp));
667 spin_lock(&info->lock);
668 if (task_tgid(current) == info->notify_owner)
669 remove_notification(info);
671 spin_unlock(&info->lock);
675 static __poll_t mqueue_poll_file(struct file *filp, struct poll_table_struct *poll_tab)
677 struct mqueue_inode_info *info = MQUEUE_I(file_inode(filp));
680 poll_wait(filp, &info->wait_q, poll_tab);
682 spin_lock(&info->lock);
683 if (info->attr.mq_curmsgs)
684 retval = EPOLLIN | EPOLLRDNORM;
686 if (info->attr.mq_curmsgs < info->attr.mq_maxmsg)
687 retval |= EPOLLOUT | EPOLLWRNORM;
688 spin_unlock(&info->lock);
693 /* Adds current to info->e_wait_q[sr] before element with smaller prio */
694 static void wq_add(struct mqueue_inode_info *info, int sr,
695 struct ext_wait_queue *ewp)
697 struct ext_wait_queue *walk;
699 list_for_each_entry(walk, &info->e_wait_q[sr].list, list) {
700 if (walk->task->prio <= current->prio) {
701 list_add_tail(&ewp->list, &walk->list);
705 list_add_tail(&ewp->list, &info->e_wait_q[sr].list);
709 * Puts current task to sleep. Caller must hold queue lock. After return
713 static int wq_sleep(struct mqueue_inode_info *info, int sr,
714 ktime_t *timeout, struct ext_wait_queue *ewp)
715 __releases(&info->lock)
720 wq_add(info, sr, ewp);
723 /* memory barrier not required, we hold info->lock */
724 __set_current_state(TASK_INTERRUPTIBLE);
726 spin_unlock(&info->lock);
727 time = schedule_hrtimeout_range_clock(timeout, 0,
728 HRTIMER_MODE_ABS, CLOCK_REALTIME);
730 if (READ_ONCE(ewp->state) == STATE_READY) {
731 /* see MQ_BARRIER for purpose/pairing */
732 smp_acquire__after_ctrl_dep();
736 spin_lock(&info->lock);
738 /* we hold info->lock, so no memory barrier required */
739 if (READ_ONCE(ewp->state) == STATE_READY) {
743 if (signal_pending(current)) {
744 retval = -ERESTARTSYS;
752 list_del(&ewp->list);
754 spin_unlock(&info->lock);
760 * Returns waiting task that should be serviced first or NULL if none exists
762 static struct ext_wait_queue *wq_get_first_waiter(
763 struct mqueue_inode_info *info, int sr)
765 struct list_head *ptr;
767 ptr = info->e_wait_q[sr].list.prev;
768 if (ptr == &info->e_wait_q[sr].list)
770 return list_entry(ptr, struct ext_wait_queue, list);
774 static inline void set_cookie(struct sk_buff *skb, char code)
776 ((char *)skb->data)[NOTIFY_COOKIE_LEN-1] = code;
780 * The next function is only to split too long sys_mq_timedsend
782 static void __do_notify(struct mqueue_inode_info *info)
785 * invoked when there is registered process and there isn't process
786 * waiting synchronously for message AND state of queue changed from
787 * empty to not empty. Here we are sure that no one is waiting
789 if (info->notify_owner &&
790 info->attr.mq_curmsgs == 1) {
791 switch (info->notify.sigev_notify) {
795 struct kernel_siginfo sig_i;
796 struct task_struct *task;
798 /* do_mq_notify() accepts sigev_signo == 0, why?? */
799 if (!info->notify.sigev_signo)
802 clear_siginfo(&sig_i);
803 sig_i.si_signo = info->notify.sigev_signo;
805 sig_i.si_code = SI_MESGQ;
806 sig_i.si_value = info->notify.sigev_value;
808 /* map current pid/uid into info->owner's namespaces */
809 sig_i.si_pid = task_tgid_nr_ns(current,
810 ns_of_pid(info->notify_owner));
811 sig_i.si_uid = from_kuid_munged(info->notify_user_ns,
814 * We can't use kill_pid_info(), this signal should
815 * bypass check_kill_permission(). It is from kernel
816 * but si_fromuser() can't know this.
817 * We do check the self_exec_id, to avoid sending
818 * signals to programs that don't expect them.
820 task = pid_task(info->notify_owner, PIDTYPE_TGID);
821 if (task && task->self_exec_id ==
822 info->notify_self_exec_id) {
823 do_send_sig_info(info->notify.sigev_signo,
824 &sig_i, task, PIDTYPE_TGID);
830 set_cookie(info->notify_cookie, NOTIFY_WOKENUP);
831 netlink_sendskb(info->notify_sock, info->notify_cookie);
834 /* after notification unregisters process */
835 put_pid(info->notify_owner);
836 put_user_ns(info->notify_user_ns);
837 info->notify_owner = NULL;
838 info->notify_user_ns = NULL;
840 wake_up(&info->wait_q);
843 static int prepare_timeout(const struct __kernel_timespec __user *u_abs_timeout,
844 struct timespec64 *ts)
846 if (get_timespec64(ts, u_abs_timeout))
848 if (!timespec64_valid(ts))
853 static void remove_notification(struct mqueue_inode_info *info)
855 if (info->notify_owner != NULL &&
856 info->notify.sigev_notify == SIGEV_THREAD) {
857 set_cookie(info->notify_cookie, NOTIFY_REMOVED);
858 netlink_sendskb(info->notify_sock, info->notify_cookie);
860 put_pid(info->notify_owner);
861 put_user_ns(info->notify_user_ns);
862 info->notify_owner = NULL;
863 info->notify_user_ns = NULL;
866 static int prepare_open(struct dentry *dentry, int oflag, int ro,
867 umode_t mode, struct filename *name,
868 struct mq_attr *attr)
870 static const int oflag2acc[O_ACCMODE] = { MAY_READ, MAY_WRITE,
871 MAY_READ | MAY_WRITE };
874 if (d_really_is_negative(dentry)) {
875 if (!(oflag & O_CREAT))
879 audit_inode_parent_hidden(name, dentry->d_parent);
880 return vfs_mkobj(dentry, mode & ~current_umask(),
881 mqueue_create_attr, attr);
883 /* it already existed */
884 audit_inode(name, dentry, 0);
885 if ((oflag & (O_CREAT|O_EXCL)) == (O_CREAT|O_EXCL))
887 if ((oflag & O_ACCMODE) == (O_RDWR | O_WRONLY))
889 acc = oflag2acc[oflag & O_ACCMODE];
890 return inode_permission(&init_user_ns, d_inode(dentry), acc);
893 static int do_mq_open(const char __user *u_name, int oflag, umode_t mode,
894 struct mq_attr *attr)
896 struct vfsmount *mnt = current->nsproxy->ipc_ns->mq_mnt;
897 struct dentry *root = mnt->mnt_root;
898 struct filename *name;
903 audit_mq_open(oflag, mode, attr);
905 if (IS_ERR(name = getname(u_name)))
906 return PTR_ERR(name);
908 fd = get_unused_fd_flags(O_CLOEXEC);
912 ro = mnt_want_write(mnt); /* we'll drop it in any case */
913 inode_lock(d_inode(root));
914 path.dentry = lookup_one_len(name->name, root, strlen(name->name));
915 if (IS_ERR(path.dentry)) {
916 error = PTR_ERR(path.dentry);
919 path.mnt = mntget(mnt);
920 error = prepare_open(path.dentry, oflag, ro, mode, name, attr);
922 struct file *file = dentry_open(&path, oflag, current_cred());
924 fd_install(fd, file);
926 error = PTR_ERR(file);
934 inode_unlock(d_inode(root));
942 SYSCALL_DEFINE4(mq_open, const char __user *, u_name, int, oflag, umode_t, mode,
943 struct mq_attr __user *, u_attr)
946 if (u_attr && copy_from_user(&attr, u_attr, sizeof(struct mq_attr)))
949 return do_mq_open(u_name, oflag, mode, u_attr ? &attr : NULL);
952 SYSCALL_DEFINE1(mq_unlink, const char __user *, u_name)
955 struct filename *name;
956 struct dentry *dentry;
957 struct inode *inode = NULL;
958 struct ipc_namespace *ipc_ns = current->nsproxy->ipc_ns;
959 struct vfsmount *mnt = ipc_ns->mq_mnt;
961 name = getname(u_name);
963 return PTR_ERR(name);
965 audit_inode_parent_hidden(name, mnt->mnt_root);
966 err = mnt_want_write(mnt);
969 inode_lock_nested(d_inode(mnt->mnt_root), I_MUTEX_PARENT);
970 dentry = lookup_one_len(name->name, mnt->mnt_root,
972 if (IS_ERR(dentry)) {
973 err = PTR_ERR(dentry);
977 inode = d_inode(dentry);
982 err = vfs_unlink(&init_user_ns, d_inode(dentry->d_parent),
988 inode_unlock(d_inode(mnt->mnt_root));
998 /* Pipelined send and receive functions.
1000 * If a receiver finds no waiting message, then it registers itself in the
1001 * list of waiting receivers. A sender checks that list before adding the new
1002 * message into the message array. If there is a waiting receiver, then it
1003 * bypasses the message array and directly hands the message over to the
1004 * receiver. The receiver accepts the message and returns without grabbing the
1007 * - Set pointer to message.
1008 * - Queue the receiver task for later wakeup (without the info->lock).
1009 * - Update its state to STATE_READY. Now the receiver can continue.
1010 * - Wake up the process after the lock is dropped. Should the process wake up
1011 * before this wakeup (due to a timeout or a signal) it will either see
1012 * STATE_READY and continue or acquire the lock to check the state again.
1014 * The same algorithm is used for senders.
1017 static inline void __pipelined_op(struct wake_q_head *wake_q,
1018 struct mqueue_inode_info *info,
1019 struct ext_wait_queue *this)
1021 struct task_struct *task;
1023 list_del(&this->list);
1024 task = get_task_struct(this->task);
1026 /* see MQ_BARRIER for purpose/pairing */
1027 smp_store_release(&this->state, STATE_READY);
1028 wake_q_add_safe(wake_q, task);
1031 /* pipelined_send() - send a message directly to the task waiting in
1032 * sys_mq_timedreceive() (without inserting message into a queue).
1034 static inline void pipelined_send(struct wake_q_head *wake_q,
1035 struct mqueue_inode_info *info,
1036 struct msg_msg *message,
1037 struct ext_wait_queue *receiver)
1039 receiver->msg = message;
1040 __pipelined_op(wake_q, info, receiver);
1043 /* pipelined_receive() - if there is task waiting in sys_mq_timedsend()
1044 * gets its message and put to the queue (we have one free place for sure). */
1045 static inline void pipelined_receive(struct wake_q_head *wake_q,
1046 struct mqueue_inode_info *info)
1048 struct ext_wait_queue *sender = wq_get_first_waiter(info, SEND);
1052 wake_up_interruptible(&info->wait_q);
1055 if (msg_insert(sender->msg, info))
1058 __pipelined_op(wake_q, info, sender);
1061 static int do_mq_timedsend(mqd_t mqdes, const char __user *u_msg_ptr,
1062 size_t msg_len, unsigned int msg_prio,
1063 struct timespec64 *ts)
1066 struct inode *inode;
1067 struct ext_wait_queue wait;
1068 struct ext_wait_queue *receiver;
1069 struct msg_msg *msg_ptr;
1070 struct mqueue_inode_info *info;
1071 ktime_t expires, *timeout = NULL;
1072 struct posix_msg_tree_node *new_leaf = NULL;
1074 DEFINE_WAKE_Q(wake_q);
1076 if (unlikely(msg_prio >= (unsigned long) MQ_PRIO_MAX))
1080 expires = timespec64_to_ktime(*ts);
1084 audit_mq_sendrecv(mqdes, msg_len, msg_prio, ts);
1087 if (unlikely(!f.file)) {
1092 inode = file_inode(f.file);
1093 if (unlikely(f.file->f_op != &mqueue_file_operations)) {
1097 info = MQUEUE_I(inode);
1100 if (unlikely(!(f.file->f_mode & FMODE_WRITE))) {
1105 if (unlikely(msg_len > info->attr.mq_msgsize)) {
1110 /* First try to allocate memory, before doing anything with
1111 * existing queues. */
1112 msg_ptr = load_msg(u_msg_ptr, msg_len);
1113 if (IS_ERR(msg_ptr)) {
1114 ret = PTR_ERR(msg_ptr);
1117 msg_ptr->m_ts = msg_len;
1118 msg_ptr->m_type = msg_prio;
1121 * msg_insert really wants us to have a valid, spare node struct so
1122 * it doesn't have to kmalloc a GFP_ATOMIC allocation, but it will
1123 * fall back to that if necessary.
1125 if (!info->node_cache)
1126 new_leaf = kmalloc(sizeof(*new_leaf), GFP_KERNEL);
1128 spin_lock(&info->lock);
1130 if (!info->node_cache && new_leaf) {
1131 /* Save our speculative allocation into the cache */
1132 INIT_LIST_HEAD(&new_leaf->msg_list);
1133 info->node_cache = new_leaf;
1139 if (info->attr.mq_curmsgs == info->attr.mq_maxmsg) {
1140 if (f.file->f_flags & O_NONBLOCK) {
1143 wait.task = current;
1144 wait.msg = (void *) msg_ptr;
1146 /* memory barrier not required, we hold info->lock */
1147 WRITE_ONCE(wait.state, STATE_NONE);
1148 ret = wq_sleep(info, SEND, timeout, &wait);
1150 * wq_sleep must be called with info->lock held, and
1151 * returns with the lock released
1156 receiver = wq_get_first_waiter(info, RECV);
1158 pipelined_send(&wake_q, info, msg_ptr, receiver);
1160 /* adds message to the queue */
1161 ret = msg_insert(msg_ptr, info);
1166 inode->i_atime = inode->i_mtime = inode->i_ctime =
1167 current_time(inode);
1170 spin_unlock(&info->lock);
1181 static int do_mq_timedreceive(mqd_t mqdes, char __user *u_msg_ptr,
1182 size_t msg_len, unsigned int __user *u_msg_prio,
1183 struct timespec64 *ts)
1186 struct msg_msg *msg_ptr;
1188 struct inode *inode;
1189 struct mqueue_inode_info *info;
1190 struct ext_wait_queue wait;
1191 ktime_t expires, *timeout = NULL;
1192 struct posix_msg_tree_node *new_leaf = NULL;
1195 expires = timespec64_to_ktime(*ts);
1199 audit_mq_sendrecv(mqdes, msg_len, 0, ts);
1202 if (unlikely(!f.file)) {
1207 inode = file_inode(f.file);
1208 if (unlikely(f.file->f_op != &mqueue_file_operations)) {
1212 info = MQUEUE_I(inode);
1215 if (unlikely(!(f.file->f_mode & FMODE_READ))) {
1220 /* checks if buffer is big enough */
1221 if (unlikely(msg_len < info->attr.mq_msgsize)) {
1227 * msg_insert really wants us to have a valid, spare node struct so
1228 * it doesn't have to kmalloc a GFP_ATOMIC allocation, but it will
1229 * fall back to that if necessary.
1231 if (!info->node_cache)
1232 new_leaf = kmalloc(sizeof(*new_leaf), GFP_KERNEL);
1234 spin_lock(&info->lock);
1236 if (!info->node_cache && new_leaf) {
1237 /* Save our speculative allocation into the cache */
1238 INIT_LIST_HEAD(&new_leaf->msg_list);
1239 info->node_cache = new_leaf;
1244 if (info->attr.mq_curmsgs == 0) {
1245 if (f.file->f_flags & O_NONBLOCK) {
1246 spin_unlock(&info->lock);
1249 wait.task = current;
1251 /* memory barrier not required, we hold info->lock */
1252 WRITE_ONCE(wait.state, STATE_NONE);
1253 ret = wq_sleep(info, RECV, timeout, &wait);
1257 DEFINE_WAKE_Q(wake_q);
1259 msg_ptr = msg_get(info);
1261 inode->i_atime = inode->i_mtime = inode->i_ctime =
1262 current_time(inode);
1264 /* There is now free space in queue. */
1265 pipelined_receive(&wake_q, info);
1266 spin_unlock(&info->lock);
1271 ret = msg_ptr->m_ts;
1273 if ((u_msg_prio && put_user(msg_ptr->m_type, u_msg_prio)) ||
1274 store_msg(u_msg_ptr, msg_ptr, msg_ptr->m_ts)) {
1285 SYSCALL_DEFINE5(mq_timedsend, mqd_t, mqdes, const char __user *, u_msg_ptr,
1286 size_t, msg_len, unsigned int, msg_prio,
1287 const struct __kernel_timespec __user *, u_abs_timeout)
1289 struct timespec64 ts, *p = NULL;
1290 if (u_abs_timeout) {
1291 int res = prepare_timeout(u_abs_timeout, &ts);
1296 return do_mq_timedsend(mqdes, u_msg_ptr, msg_len, msg_prio, p);
1299 SYSCALL_DEFINE5(mq_timedreceive, mqd_t, mqdes, char __user *, u_msg_ptr,
1300 size_t, msg_len, unsigned int __user *, u_msg_prio,
1301 const struct __kernel_timespec __user *, u_abs_timeout)
1303 struct timespec64 ts, *p = NULL;
1304 if (u_abs_timeout) {
1305 int res = prepare_timeout(u_abs_timeout, &ts);
1310 return do_mq_timedreceive(mqdes, u_msg_ptr, msg_len, u_msg_prio, p);
1314 * Notes: the case when user wants us to deregister (with NULL as pointer)
1315 * and he isn't currently owner of notification, will be silently discarded.
1316 * It isn't explicitly defined in the POSIX.
1318 static int do_mq_notify(mqd_t mqdes, const struct sigevent *notification)
1323 struct inode *inode;
1324 struct mqueue_inode_info *info;
1327 audit_mq_notify(mqdes, notification);
1331 if (notification != NULL) {
1332 if (unlikely(notification->sigev_notify != SIGEV_NONE &&
1333 notification->sigev_notify != SIGEV_SIGNAL &&
1334 notification->sigev_notify != SIGEV_THREAD))
1336 if (notification->sigev_notify == SIGEV_SIGNAL &&
1337 !valid_signal(notification->sigev_signo)) {
1340 if (notification->sigev_notify == SIGEV_THREAD) {
1343 /* create the notify skb */
1344 nc = alloc_skb(NOTIFY_COOKIE_LEN, GFP_KERNEL);
1348 if (copy_from_user(nc->data,
1349 notification->sigev_value.sival_ptr,
1350 NOTIFY_COOKIE_LEN)) {
1355 /* TODO: add a header? */
1356 skb_put(nc, NOTIFY_COOKIE_LEN);
1357 /* and attach it to the socket */
1359 f = fdget(notification->sigev_signo);
1364 sock = netlink_getsockbyfilp(f.file);
1367 ret = PTR_ERR(sock);
1371 timeo = MAX_SCHEDULE_TIMEOUT;
1372 ret = netlink_attachskb(sock, nc, &timeo, NULL);
1388 inode = file_inode(f.file);
1389 if (unlikely(f.file->f_op != &mqueue_file_operations)) {
1393 info = MQUEUE_I(inode);
1396 spin_lock(&info->lock);
1397 if (notification == NULL) {
1398 if (info->notify_owner == task_tgid(current)) {
1399 remove_notification(info);
1400 inode->i_atime = inode->i_ctime = current_time(inode);
1402 } else if (info->notify_owner != NULL) {
1405 switch (notification->sigev_notify) {
1407 info->notify.sigev_notify = SIGEV_NONE;
1410 info->notify_sock = sock;
1411 info->notify_cookie = nc;
1414 info->notify.sigev_notify = SIGEV_THREAD;
1417 info->notify.sigev_signo = notification->sigev_signo;
1418 info->notify.sigev_value = notification->sigev_value;
1419 info->notify.sigev_notify = SIGEV_SIGNAL;
1420 info->notify_self_exec_id = current->self_exec_id;
1424 info->notify_owner = get_pid(task_tgid(current));
1425 info->notify_user_ns = get_user_ns(current_user_ns());
1426 inode->i_atime = inode->i_ctime = current_time(inode);
1428 spin_unlock(&info->lock);
1433 netlink_detachskb(sock, nc);
1441 SYSCALL_DEFINE2(mq_notify, mqd_t, mqdes,
1442 const struct sigevent __user *, u_notification)
1444 struct sigevent n, *p = NULL;
1445 if (u_notification) {
1446 if (copy_from_user(&n, u_notification, sizeof(struct sigevent)))
1450 return do_mq_notify(mqdes, p);
1453 static int do_mq_getsetattr(int mqdes, struct mq_attr *new, struct mq_attr *old)
1456 struct inode *inode;
1457 struct mqueue_inode_info *info;
1459 if (new && (new->mq_flags & (~O_NONBLOCK)))
1466 if (unlikely(f.file->f_op != &mqueue_file_operations)) {
1471 inode = file_inode(f.file);
1472 info = MQUEUE_I(inode);
1474 spin_lock(&info->lock);
1478 old->mq_flags = f.file->f_flags & O_NONBLOCK;
1481 audit_mq_getsetattr(mqdes, new);
1482 spin_lock(&f.file->f_lock);
1483 if (new->mq_flags & O_NONBLOCK)
1484 f.file->f_flags |= O_NONBLOCK;
1486 f.file->f_flags &= ~O_NONBLOCK;
1487 spin_unlock(&f.file->f_lock);
1489 inode->i_atime = inode->i_ctime = current_time(inode);
1492 spin_unlock(&info->lock);
1497 SYSCALL_DEFINE3(mq_getsetattr, mqd_t, mqdes,
1498 const struct mq_attr __user *, u_mqstat,
1499 struct mq_attr __user *, u_omqstat)
1502 struct mq_attr mqstat, omqstat;
1503 struct mq_attr *new = NULL, *old = NULL;
1507 if (copy_from_user(new, u_mqstat, sizeof(struct mq_attr)))
1513 ret = do_mq_getsetattr(mqdes, new, old);
1517 if (copy_to_user(u_omqstat, old, sizeof(struct mq_attr)))
1522 #ifdef CONFIG_COMPAT
1524 struct compat_mq_attr {
1525 compat_long_t mq_flags; /* message queue flags */
1526 compat_long_t mq_maxmsg; /* maximum number of messages */
1527 compat_long_t mq_msgsize; /* maximum message size */
1528 compat_long_t mq_curmsgs; /* number of messages currently queued */
1529 compat_long_t __reserved[4]; /* ignored for input, zeroed for output */
1532 static inline int get_compat_mq_attr(struct mq_attr *attr,
1533 const struct compat_mq_attr __user *uattr)
1535 struct compat_mq_attr v;
1537 if (copy_from_user(&v, uattr, sizeof(*uattr)))
1540 memset(attr, 0, sizeof(*attr));
1541 attr->mq_flags = v.mq_flags;
1542 attr->mq_maxmsg = v.mq_maxmsg;
1543 attr->mq_msgsize = v.mq_msgsize;
1544 attr->mq_curmsgs = v.mq_curmsgs;
1548 static inline int put_compat_mq_attr(const struct mq_attr *attr,
1549 struct compat_mq_attr __user *uattr)
1551 struct compat_mq_attr v;
1553 memset(&v, 0, sizeof(v));
1554 v.mq_flags = attr->mq_flags;
1555 v.mq_maxmsg = attr->mq_maxmsg;
1556 v.mq_msgsize = attr->mq_msgsize;
1557 v.mq_curmsgs = attr->mq_curmsgs;
1558 if (copy_to_user(uattr, &v, sizeof(*uattr)))
1563 COMPAT_SYSCALL_DEFINE4(mq_open, const char __user *, u_name,
1564 int, oflag, compat_mode_t, mode,
1565 struct compat_mq_attr __user *, u_attr)
1567 struct mq_attr attr, *p = NULL;
1568 if (u_attr && oflag & O_CREAT) {
1570 if (get_compat_mq_attr(&attr, u_attr))
1573 return do_mq_open(u_name, oflag, mode, p);
1576 COMPAT_SYSCALL_DEFINE2(mq_notify, mqd_t, mqdes,
1577 const struct compat_sigevent __user *, u_notification)
1579 struct sigevent n, *p = NULL;
1580 if (u_notification) {
1581 if (get_compat_sigevent(&n, u_notification))
1583 if (n.sigev_notify == SIGEV_THREAD)
1584 n.sigev_value.sival_ptr = compat_ptr(n.sigev_value.sival_int);
1587 return do_mq_notify(mqdes, p);
1590 COMPAT_SYSCALL_DEFINE3(mq_getsetattr, mqd_t, mqdes,
1591 const struct compat_mq_attr __user *, u_mqstat,
1592 struct compat_mq_attr __user *, u_omqstat)
1595 struct mq_attr mqstat, omqstat;
1596 struct mq_attr *new = NULL, *old = NULL;
1600 if (get_compat_mq_attr(new, u_mqstat))
1606 ret = do_mq_getsetattr(mqdes, new, old);
1610 if (put_compat_mq_attr(old, u_omqstat))
1616 #ifdef CONFIG_COMPAT_32BIT_TIME
1617 static int compat_prepare_timeout(const struct old_timespec32 __user *p,
1618 struct timespec64 *ts)
1620 if (get_old_timespec32(ts, p))
1622 if (!timespec64_valid(ts))
1627 SYSCALL_DEFINE5(mq_timedsend_time32, mqd_t, mqdes,
1628 const char __user *, u_msg_ptr,
1629 unsigned int, msg_len, unsigned int, msg_prio,
1630 const struct old_timespec32 __user *, u_abs_timeout)
1632 struct timespec64 ts, *p = NULL;
1633 if (u_abs_timeout) {
1634 int res = compat_prepare_timeout(u_abs_timeout, &ts);
1639 return do_mq_timedsend(mqdes, u_msg_ptr, msg_len, msg_prio, p);
1642 SYSCALL_DEFINE5(mq_timedreceive_time32, mqd_t, mqdes,
1643 char __user *, u_msg_ptr,
1644 unsigned int, msg_len, unsigned int __user *, u_msg_prio,
1645 const struct old_timespec32 __user *, u_abs_timeout)
1647 struct timespec64 ts, *p = NULL;
1648 if (u_abs_timeout) {
1649 int res = compat_prepare_timeout(u_abs_timeout, &ts);
1654 return do_mq_timedreceive(mqdes, u_msg_ptr, msg_len, u_msg_prio, p);
1658 static const struct inode_operations mqueue_dir_inode_operations = {
1659 .lookup = simple_lookup,
1660 .create = mqueue_create,
1661 .unlink = mqueue_unlink,
1664 static const struct file_operations mqueue_file_operations = {
1665 .flush = mqueue_flush_file,
1666 .poll = mqueue_poll_file,
1667 .read = mqueue_read_file,
1668 .llseek = default_llseek,
1671 static const struct super_operations mqueue_super_ops = {
1672 .alloc_inode = mqueue_alloc_inode,
1673 .free_inode = mqueue_free_inode,
1674 .evict_inode = mqueue_evict_inode,
1675 .statfs = simple_statfs,
1678 static const struct fs_context_operations mqueue_fs_context_ops = {
1679 .free = mqueue_fs_context_free,
1680 .get_tree = mqueue_get_tree,
1683 static struct file_system_type mqueue_fs_type = {
1685 .init_fs_context = mqueue_init_fs_context,
1686 .kill_sb = kill_litter_super,
1687 .fs_flags = FS_USERNS_MOUNT,
1690 int mq_init_ns(struct ipc_namespace *ns)
1694 ns->mq_queues_count = 0;
1695 ns->mq_queues_max = DFLT_QUEUESMAX;
1696 ns->mq_msg_max = DFLT_MSGMAX;
1697 ns->mq_msgsize_max = DFLT_MSGSIZEMAX;
1698 ns->mq_msg_default = DFLT_MSG;
1699 ns->mq_msgsize_default = DFLT_MSGSIZE;
1701 m = mq_create_mount(ns);
1708 void mq_clear_sbinfo(struct ipc_namespace *ns)
1710 ns->mq_mnt->mnt_sb->s_fs_info = NULL;
1713 void mq_put_mnt(struct ipc_namespace *ns)
1715 kern_unmount(ns->mq_mnt);
1718 static int __init init_mqueue_fs(void)
1722 mqueue_inode_cachep = kmem_cache_create("mqueue_inode_cache",
1723 sizeof(struct mqueue_inode_info), 0,
1724 SLAB_HWCACHE_ALIGN|SLAB_ACCOUNT, init_once);
1725 if (mqueue_inode_cachep == NULL)
1728 if (!setup_mq_sysctls(&init_ipc_ns)) {
1729 pr_warn("sysctl registration failed\n");
1733 error = register_filesystem(&mqueue_fs_type);
1737 spin_lock_init(&mq_lock);
1739 error = mq_init_ns(&init_ipc_ns);
1741 goto out_filesystem;
1746 unregister_filesystem(&mqueue_fs_type);
1748 kmem_cache_destroy(mqueue_inode_cachep);
1749 retire_mq_sysctls(&init_ipc_ns);
1753 device_initcall(init_mqueue_fs);