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;
50 #define MQUEUE_MAGIC 0x19800202
51 #define DIRENT_SIZE 20
52 #define FILENT_SIZE 80
60 struct posix_msg_tree_node {
61 struct rb_node rb_node;
62 struct list_head msg_list;
69 * Accesses to a message queue are synchronized by acquiring info->lock.
71 * There are two notable exceptions:
72 * - The actual wakeup of a sleeping task is performed using the wake_q
73 * framework. info->lock is already released when wake_up_q is called.
74 * - The exit codepaths after sleeping check ext_wait_queue->state without
75 * any locks. If it is STATE_READY, then the syscall is completed without
76 * acquiring info->lock.
79 * To achieve proper release/acquire memory barrier pairing, the state is set to
80 * STATE_READY with smp_store_release(), and it is read with READ_ONCE followed
81 * by smp_acquire__after_ctrl_dep(). In addition, wake_q_add_safe() is used.
83 * This prevents the following races:
85 * 1) With the simple wake_q_add(), the task could be gone already before
86 * the increase of the reference happens
89 * WRITE_ONCE(wait.state, STATE_NONE);
90 * schedule_hrtimeout()
92 * if (cmpxchg()) // success
93 * ->state = STATE_READY (reordered)
95 * if (wait.state == STATE_READY) return;
96 * sysret to user space
98 * get_task_struct() // UaF
100 * Solution: Use wake_q_add_safe() and perform the get_task_struct() before
101 * the smp_store_release() that does ->state = STATE_READY.
103 * 2) Without proper _release/_acquire barriers, the woken up task
104 * could read stale data
109 * WRITE_ONCE(wait.state, STATE_NONE);
110 * schedule_hrtimeout()
111 * state = STATE_READY;
113 * if (wait.state == STATE_READY) return;
114 * msg_ptr = wait.msg; // Access to stale data!
115 * receiver->msg = message; (reordered)
117 * Solution: use _release and _acquire barriers.
119 * 3) There is intentionally no barrier when setting current->state
120 * to TASK_INTERRUPTIBLE: spin_unlock(&info->lock) provides the
121 * release memory barrier, and the wakeup is triggered when holding
122 * info->lock, i.e. spin_lock(&info->lock) provided a pairing
123 * acquire memory barrier.
126 struct ext_wait_queue { /* queue of sleeping tasks */
127 struct task_struct *task;
128 struct list_head list;
129 struct msg_msg *msg; /* ptr of loaded message */
130 int state; /* one of STATE_* values */
133 struct mqueue_inode_info {
135 struct inode vfs_inode;
136 wait_queue_head_t wait_q;
138 struct rb_root msg_tree;
139 struct rb_node *msg_tree_rightmost;
140 struct posix_msg_tree_node *node_cache;
143 struct sigevent notify;
144 struct pid *notify_owner;
145 u32 notify_self_exec_id;
146 struct user_namespace *notify_user_ns;
147 struct ucounts *ucounts; /* user who created, for accounting */
148 struct sock *notify_sock;
149 struct sk_buff *notify_cookie;
151 /* for tasks waiting for free space and messages, respectively */
152 struct ext_wait_queue e_wait_q[2];
154 unsigned long qsize; /* size of queue in memory (sum of all msgs) */
157 static struct file_system_type mqueue_fs_type;
158 static const struct inode_operations mqueue_dir_inode_operations;
159 static const struct file_operations mqueue_file_operations;
160 static const struct super_operations mqueue_super_ops;
161 static const struct fs_context_operations mqueue_fs_context_ops;
162 static void remove_notification(struct mqueue_inode_info *info);
164 static struct kmem_cache *mqueue_inode_cachep;
166 static struct ctl_table_header *mq_sysctl_table;
168 static inline struct mqueue_inode_info *MQUEUE_I(struct inode *inode)
170 return container_of(inode, struct mqueue_inode_info, vfs_inode);
174 * This routine should be called with the mq_lock held.
176 static inline struct ipc_namespace *__get_ns_from_inode(struct inode *inode)
178 return get_ipc_ns(inode->i_sb->s_fs_info);
181 static struct ipc_namespace *get_ns_from_inode(struct inode *inode)
183 struct ipc_namespace *ns;
186 ns = __get_ns_from_inode(inode);
187 spin_unlock(&mq_lock);
191 /* Auxiliary functions to manipulate messages' list */
192 static int msg_insert(struct msg_msg *msg, struct mqueue_inode_info *info)
194 struct rb_node **p, *parent = NULL;
195 struct posix_msg_tree_node *leaf;
196 bool rightmost = true;
198 p = &info->msg_tree.rb_node;
201 leaf = rb_entry(parent, struct posix_msg_tree_node, rb_node);
203 if (likely(leaf->priority == msg->m_type))
205 else if (msg->m_type < leaf->priority) {
211 if (info->node_cache) {
212 leaf = info->node_cache;
213 info->node_cache = NULL;
215 leaf = kmalloc(sizeof(*leaf), GFP_ATOMIC);
218 INIT_LIST_HEAD(&leaf->msg_list);
220 leaf->priority = msg->m_type;
223 info->msg_tree_rightmost = &leaf->rb_node;
225 rb_link_node(&leaf->rb_node, parent, p);
226 rb_insert_color(&leaf->rb_node, &info->msg_tree);
228 info->attr.mq_curmsgs++;
229 info->qsize += msg->m_ts;
230 list_add_tail(&msg->m_list, &leaf->msg_list);
234 static inline void msg_tree_erase(struct posix_msg_tree_node *leaf,
235 struct mqueue_inode_info *info)
237 struct rb_node *node = &leaf->rb_node;
239 if (info->msg_tree_rightmost == node)
240 info->msg_tree_rightmost = rb_prev(node);
242 rb_erase(node, &info->msg_tree);
243 if (info->node_cache)
246 info->node_cache = leaf;
249 static inline struct msg_msg *msg_get(struct mqueue_inode_info *info)
251 struct rb_node *parent = NULL;
252 struct posix_msg_tree_node *leaf;
257 * During insert, low priorities go to the left and high to the
258 * right. On receive, we want the highest priorities first, so
259 * walk all the way to the right.
261 parent = info->msg_tree_rightmost;
263 if (info->attr.mq_curmsgs) {
264 pr_warn_once("Inconsistency in POSIX message queue, "
265 "no tree element, but supposedly messages "
267 info->attr.mq_curmsgs = 0;
271 leaf = rb_entry(parent, struct posix_msg_tree_node, rb_node);
272 if (unlikely(list_empty(&leaf->msg_list))) {
273 pr_warn_once("Inconsistency in POSIX message queue, "
274 "empty leaf node but we haven't implemented "
275 "lazy leaf delete!\n");
276 msg_tree_erase(leaf, info);
279 msg = list_first_entry(&leaf->msg_list,
280 struct msg_msg, m_list);
281 list_del(&msg->m_list);
282 if (list_empty(&leaf->msg_list)) {
283 msg_tree_erase(leaf, info);
286 info->attr.mq_curmsgs--;
287 info->qsize -= msg->m_ts;
291 static struct inode *mqueue_get_inode(struct super_block *sb,
292 struct ipc_namespace *ipc_ns, umode_t mode,
293 struct mq_attr *attr)
298 inode = new_inode(sb);
302 inode->i_ino = get_next_ino();
303 inode->i_mode = mode;
304 inode->i_uid = current_fsuid();
305 inode->i_gid = current_fsgid();
306 inode->i_mtime = inode->i_ctime = inode->i_atime = current_time(inode);
309 struct mqueue_inode_info *info;
310 unsigned long mq_bytes, mq_treesize;
312 inode->i_fop = &mqueue_file_operations;
313 inode->i_size = FILENT_SIZE;
314 /* mqueue specific info */
315 info = MQUEUE_I(inode);
316 spin_lock_init(&info->lock);
317 init_waitqueue_head(&info->wait_q);
318 INIT_LIST_HEAD(&info->e_wait_q[0].list);
319 INIT_LIST_HEAD(&info->e_wait_q[1].list);
320 info->notify_owner = NULL;
321 info->notify_user_ns = NULL;
323 info->ucounts = NULL; /* set when all is ok */
324 info->msg_tree = RB_ROOT;
325 info->msg_tree_rightmost = NULL;
326 info->node_cache = NULL;
327 memset(&info->attr, 0, sizeof(info->attr));
328 info->attr.mq_maxmsg = min(ipc_ns->mq_msg_max,
329 ipc_ns->mq_msg_default);
330 info->attr.mq_msgsize = min(ipc_ns->mq_msgsize_max,
331 ipc_ns->mq_msgsize_default);
333 info->attr.mq_maxmsg = attr->mq_maxmsg;
334 info->attr.mq_msgsize = attr->mq_msgsize;
337 * We used to allocate a static array of pointers and account
338 * the size of that array as well as one msg_msg struct per
339 * possible message into the queue size. That's no longer
340 * accurate as the queue is now an rbtree and will grow and
341 * shrink depending on usage patterns. We can, however, still
342 * account one msg_msg struct per message, but the nodes are
343 * allocated depending on priority usage, and most programs
344 * only use one, or a handful, of priorities. However, since
345 * this is pinned memory, we need to assume worst case, so
346 * that means the min(mq_maxmsg, max_priorities) * struct
347 * posix_msg_tree_node.
351 if (info->attr.mq_maxmsg <= 0 || info->attr.mq_msgsize <= 0)
353 if (capable(CAP_SYS_RESOURCE)) {
354 if (info->attr.mq_maxmsg > HARD_MSGMAX ||
355 info->attr.mq_msgsize > HARD_MSGSIZEMAX)
358 if (info->attr.mq_maxmsg > ipc_ns->mq_msg_max ||
359 info->attr.mq_msgsize > ipc_ns->mq_msgsize_max)
363 /* check for overflow */
364 if (info->attr.mq_msgsize > ULONG_MAX/info->attr.mq_maxmsg)
366 mq_treesize = info->attr.mq_maxmsg * sizeof(struct msg_msg) +
367 min_t(unsigned int, info->attr.mq_maxmsg, MQ_PRIO_MAX) *
368 sizeof(struct posix_msg_tree_node);
369 mq_bytes = info->attr.mq_maxmsg * info->attr.mq_msgsize;
370 if (mq_bytes + mq_treesize < mq_bytes)
372 mq_bytes += mq_treesize;
373 info->ucounts = get_ucounts(current_ucounts());
378 msgqueue = inc_rlimit_ucounts(info->ucounts, UCOUNT_RLIMIT_MSGQUEUE, mq_bytes);
379 if (msgqueue == LONG_MAX || msgqueue > rlimit(RLIMIT_MSGQUEUE)) {
380 dec_rlimit_ucounts(info->ucounts, UCOUNT_RLIMIT_MSGQUEUE, mq_bytes);
381 spin_unlock(&mq_lock);
382 put_ucounts(info->ucounts);
383 info->ucounts = NULL;
384 /* mqueue_evict_inode() releases info->messages */
388 spin_unlock(&mq_lock);
390 } else if (S_ISDIR(mode)) {
392 /* Some things misbehave if size == 0 on a directory */
393 inode->i_size = 2 * DIRENT_SIZE;
394 inode->i_op = &mqueue_dir_inode_operations;
395 inode->i_fop = &simple_dir_operations;
405 static int mqueue_fill_super(struct super_block *sb, struct fs_context *fc)
408 struct ipc_namespace *ns = sb->s_fs_info;
410 sb->s_iflags |= SB_I_NOEXEC | SB_I_NODEV;
411 sb->s_blocksize = PAGE_SIZE;
412 sb->s_blocksize_bits = PAGE_SHIFT;
413 sb->s_magic = MQUEUE_MAGIC;
414 sb->s_op = &mqueue_super_ops;
416 inode = mqueue_get_inode(sb, ns, S_IFDIR | S_ISVTX | S_IRWXUGO, NULL);
418 return PTR_ERR(inode);
420 sb->s_root = d_make_root(inode);
426 static int mqueue_get_tree(struct fs_context *fc)
428 struct mqueue_fs_context *ctx = fc->fs_private;
430 return get_tree_keyed(fc, mqueue_fill_super, ctx->ipc_ns);
433 static void mqueue_fs_context_free(struct fs_context *fc)
435 struct mqueue_fs_context *ctx = fc->fs_private;
437 put_ipc_ns(ctx->ipc_ns);
441 static int mqueue_init_fs_context(struct fs_context *fc)
443 struct mqueue_fs_context *ctx;
445 ctx = kzalloc(sizeof(struct mqueue_fs_context), GFP_KERNEL);
449 ctx->ipc_ns = get_ipc_ns(current->nsproxy->ipc_ns);
450 put_user_ns(fc->user_ns);
451 fc->user_ns = get_user_ns(ctx->ipc_ns->user_ns);
452 fc->fs_private = ctx;
453 fc->ops = &mqueue_fs_context_ops;
457 static struct vfsmount *mq_create_mount(struct ipc_namespace *ns)
459 struct mqueue_fs_context *ctx;
460 struct fs_context *fc;
461 struct vfsmount *mnt;
463 fc = fs_context_for_mount(&mqueue_fs_type, SB_KERNMOUNT);
467 ctx = fc->fs_private;
468 put_ipc_ns(ctx->ipc_ns);
469 ctx->ipc_ns = get_ipc_ns(ns);
470 put_user_ns(fc->user_ns);
471 fc->user_ns = get_user_ns(ctx->ipc_ns->user_ns);
478 static void init_once(void *foo)
480 struct mqueue_inode_info *p = (struct mqueue_inode_info *) foo;
482 inode_init_once(&p->vfs_inode);
485 static struct inode *mqueue_alloc_inode(struct super_block *sb)
487 struct mqueue_inode_info *ei;
489 ei = alloc_inode_sb(sb, mqueue_inode_cachep, GFP_KERNEL);
492 return &ei->vfs_inode;
495 static void mqueue_free_inode(struct inode *inode)
497 kmem_cache_free(mqueue_inode_cachep, MQUEUE_I(inode));
500 static void mqueue_evict_inode(struct inode *inode)
502 struct mqueue_inode_info *info;
503 struct ipc_namespace *ipc_ns;
504 struct msg_msg *msg, *nmsg;
509 if (S_ISDIR(inode->i_mode))
512 ipc_ns = get_ns_from_inode(inode);
513 info = MQUEUE_I(inode);
514 spin_lock(&info->lock);
515 while ((msg = msg_get(info)) != NULL)
516 list_add_tail(&msg->m_list, &tmp_msg);
517 kfree(info->node_cache);
518 spin_unlock(&info->lock);
520 list_for_each_entry_safe(msg, nmsg, &tmp_msg, m_list) {
521 list_del(&msg->m_list);
526 unsigned long mq_bytes, mq_treesize;
528 /* Total amount of bytes accounted for the mqueue */
529 mq_treesize = info->attr.mq_maxmsg * sizeof(struct msg_msg) +
530 min_t(unsigned int, info->attr.mq_maxmsg, MQ_PRIO_MAX) *
531 sizeof(struct posix_msg_tree_node);
533 mq_bytes = mq_treesize + (info->attr.mq_maxmsg *
534 info->attr.mq_msgsize);
537 dec_rlimit_ucounts(info->ucounts, UCOUNT_RLIMIT_MSGQUEUE, mq_bytes);
539 * get_ns_from_inode() ensures that the
540 * (ipc_ns = sb->s_fs_info) is either a valid ipc_ns
541 * to which we now hold a reference, or it is NULL.
542 * We can't put it here under mq_lock, though.
545 ipc_ns->mq_queues_count--;
546 spin_unlock(&mq_lock);
547 put_ucounts(info->ucounts);
548 info->ucounts = NULL;
554 static int mqueue_create_attr(struct dentry *dentry, umode_t mode, void *arg)
556 struct inode *dir = dentry->d_parent->d_inode;
558 struct mq_attr *attr = arg;
560 struct ipc_namespace *ipc_ns;
563 ipc_ns = __get_ns_from_inode(dir);
569 if (ipc_ns->mq_queues_count >= ipc_ns->mq_queues_max &&
570 !capable(CAP_SYS_RESOURCE)) {
574 ipc_ns->mq_queues_count++;
575 spin_unlock(&mq_lock);
577 inode = mqueue_get_inode(dir->i_sb, ipc_ns, mode, attr);
579 error = PTR_ERR(inode);
581 ipc_ns->mq_queues_count--;
586 dir->i_size += DIRENT_SIZE;
587 dir->i_ctime = dir->i_mtime = dir->i_atime = current_time(dir);
589 d_instantiate(dentry, inode);
593 spin_unlock(&mq_lock);
599 static int mqueue_create(struct user_namespace *mnt_userns, struct inode *dir,
600 struct dentry *dentry, umode_t mode, bool excl)
602 return mqueue_create_attr(dentry, mode, NULL);
605 static int mqueue_unlink(struct inode *dir, struct dentry *dentry)
607 struct inode *inode = d_inode(dentry);
609 dir->i_ctime = dir->i_mtime = dir->i_atime = current_time(dir);
610 dir->i_size -= DIRENT_SIZE;
617 * This is routine for system read from queue file.
618 * To avoid mess with doing here some sort of mq_receive we allow
619 * to read only queue size & notification info (the only values
620 * that are interesting from user point of view and aren't accessible
621 * through std routines)
623 static ssize_t mqueue_read_file(struct file *filp, char __user *u_data,
624 size_t count, loff_t *off)
626 struct mqueue_inode_info *info = MQUEUE_I(file_inode(filp));
627 char buffer[FILENT_SIZE];
630 spin_lock(&info->lock);
631 snprintf(buffer, sizeof(buffer),
632 "QSIZE:%-10lu NOTIFY:%-5d SIGNO:%-5d NOTIFY_PID:%-6d\n",
634 info->notify_owner ? info->notify.sigev_notify : 0,
635 (info->notify_owner &&
636 info->notify.sigev_notify == SIGEV_SIGNAL) ?
637 info->notify.sigev_signo : 0,
638 pid_vnr(info->notify_owner));
639 spin_unlock(&info->lock);
640 buffer[sizeof(buffer)-1] = '\0';
642 ret = simple_read_from_buffer(u_data, count, off, buffer,
647 file_inode(filp)->i_atime = file_inode(filp)->i_ctime = current_time(file_inode(filp));
651 static int mqueue_flush_file(struct file *filp, fl_owner_t id)
653 struct mqueue_inode_info *info = MQUEUE_I(file_inode(filp));
655 spin_lock(&info->lock);
656 if (task_tgid(current) == info->notify_owner)
657 remove_notification(info);
659 spin_unlock(&info->lock);
663 static __poll_t mqueue_poll_file(struct file *filp, struct poll_table_struct *poll_tab)
665 struct mqueue_inode_info *info = MQUEUE_I(file_inode(filp));
668 poll_wait(filp, &info->wait_q, poll_tab);
670 spin_lock(&info->lock);
671 if (info->attr.mq_curmsgs)
672 retval = EPOLLIN | EPOLLRDNORM;
674 if (info->attr.mq_curmsgs < info->attr.mq_maxmsg)
675 retval |= EPOLLOUT | EPOLLWRNORM;
676 spin_unlock(&info->lock);
681 /* Adds current to info->e_wait_q[sr] before element with smaller prio */
682 static void wq_add(struct mqueue_inode_info *info, int sr,
683 struct ext_wait_queue *ewp)
685 struct ext_wait_queue *walk;
687 list_for_each_entry(walk, &info->e_wait_q[sr].list, list) {
688 if (walk->task->prio <= current->prio) {
689 list_add_tail(&ewp->list, &walk->list);
693 list_add_tail(&ewp->list, &info->e_wait_q[sr].list);
697 * Puts current task to sleep. Caller must hold queue lock. After return
701 static int wq_sleep(struct mqueue_inode_info *info, int sr,
702 ktime_t *timeout, struct ext_wait_queue *ewp)
703 __releases(&info->lock)
708 wq_add(info, sr, ewp);
711 /* memory barrier not required, we hold info->lock */
712 __set_current_state(TASK_INTERRUPTIBLE);
714 spin_unlock(&info->lock);
715 time = schedule_hrtimeout_range_clock(timeout, 0,
716 HRTIMER_MODE_ABS, CLOCK_REALTIME);
718 if (READ_ONCE(ewp->state) == STATE_READY) {
719 /* see MQ_BARRIER for purpose/pairing */
720 smp_acquire__after_ctrl_dep();
724 spin_lock(&info->lock);
726 /* we hold info->lock, so no memory barrier required */
727 if (READ_ONCE(ewp->state) == STATE_READY) {
731 if (signal_pending(current)) {
732 retval = -ERESTARTSYS;
740 list_del(&ewp->list);
742 spin_unlock(&info->lock);
748 * Returns waiting task that should be serviced first or NULL if none exists
750 static struct ext_wait_queue *wq_get_first_waiter(
751 struct mqueue_inode_info *info, int sr)
753 struct list_head *ptr;
755 ptr = info->e_wait_q[sr].list.prev;
756 if (ptr == &info->e_wait_q[sr].list)
758 return list_entry(ptr, struct ext_wait_queue, list);
762 static inline void set_cookie(struct sk_buff *skb, char code)
764 ((char *)skb->data)[NOTIFY_COOKIE_LEN-1] = code;
768 * The next function is only to split too long sys_mq_timedsend
770 static void __do_notify(struct mqueue_inode_info *info)
773 * invoked when there is registered process and there isn't process
774 * waiting synchronously for message AND state of queue changed from
775 * empty to not empty. Here we are sure that no one is waiting
777 if (info->notify_owner &&
778 info->attr.mq_curmsgs == 1) {
779 switch (info->notify.sigev_notify) {
783 struct kernel_siginfo sig_i;
784 struct task_struct *task;
786 /* do_mq_notify() accepts sigev_signo == 0, why?? */
787 if (!info->notify.sigev_signo)
790 clear_siginfo(&sig_i);
791 sig_i.si_signo = info->notify.sigev_signo;
793 sig_i.si_code = SI_MESGQ;
794 sig_i.si_value = info->notify.sigev_value;
796 /* map current pid/uid into info->owner's namespaces */
797 sig_i.si_pid = task_tgid_nr_ns(current,
798 ns_of_pid(info->notify_owner));
799 sig_i.si_uid = from_kuid_munged(info->notify_user_ns,
802 * We can't use kill_pid_info(), this signal should
803 * bypass check_kill_permission(). It is from kernel
804 * but si_fromuser() can't know this.
805 * We do check the self_exec_id, to avoid sending
806 * signals to programs that don't expect them.
808 task = pid_task(info->notify_owner, PIDTYPE_TGID);
809 if (task && task->self_exec_id ==
810 info->notify_self_exec_id) {
811 do_send_sig_info(info->notify.sigev_signo,
812 &sig_i, task, PIDTYPE_TGID);
818 set_cookie(info->notify_cookie, NOTIFY_WOKENUP);
819 netlink_sendskb(info->notify_sock, info->notify_cookie);
822 /* after notification unregisters process */
823 put_pid(info->notify_owner);
824 put_user_ns(info->notify_user_ns);
825 info->notify_owner = NULL;
826 info->notify_user_ns = NULL;
828 wake_up(&info->wait_q);
831 static int prepare_timeout(const struct __kernel_timespec __user *u_abs_timeout,
832 struct timespec64 *ts)
834 if (get_timespec64(ts, u_abs_timeout))
836 if (!timespec64_valid(ts))
841 static void remove_notification(struct mqueue_inode_info *info)
843 if (info->notify_owner != NULL &&
844 info->notify.sigev_notify == SIGEV_THREAD) {
845 set_cookie(info->notify_cookie, NOTIFY_REMOVED);
846 netlink_sendskb(info->notify_sock, info->notify_cookie);
848 put_pid(info->notify_owner);
849 put_user_ns(info->notify_user_ns);
850 info->notify_owner = NULL;
851 info->notify_user_ns = NULL;
854 static int prepare_open(struct dentry *dentry, int oflag, int ro,
855 umode_t mode, struct filename *name,
856 struct mq_attr *attr)
858 static const int oflag2acc[O_ACCMODE] = { MAY_READ, MAY_WRITE,
859 MAY_READ | MAY_WRITE };
862 if (d_really_is_negative(dentry)) {
863 if (!(oflag & O_CREAT))
867 audit_inode_parent_hidden(name, dentry->d_parent);
868 return vfs_mkobj(dentry, mode & ~current_umask(),
869 mqueue_create_attr, attr);
871 /* it already existed */
872 audit_inode(name, dentry, 0);
873 if ((oflag & (O_CREAT|O_EXCL)) == (O_CREAT|O_EXCL))
875 if ((oflag & O_ACCMODE) == (O_RDWR | O_WRONLY))
877 acc = oflag2acc[oflag & O_ACCMODE];
878 return inode_permission(&init_user_ns, d_inode(dentry), acc);
881 static int do_mq_open(const char __user *u_name, int oflag, umode_t mode,
882 struct mq_attr *attr)
884 struct vfsmount *mnt = current->nsproxy->ipc_ns->mq_mnt;
885 struct dentry *root = mnt->mnt_root;
886 struct filename *name;
891 audit_mq_open(oflag, mode, attr);
893 if (IS_ERR(name = getname(u_name)))
894 return PTR_ERR(name);
896 fd = get_unused_fd_flags(O_CLOEXEC);
900 ro = mnt_want_write(mnt); /* we'll drop it in any case */
901 inode_lock(d_inode(root));
902 path.dentry = lookup_one_len(name->name, root, strlen(name->name));
903 if (IS_ERR(path.dentry)) {
904 error = PTR_ERR(path.dentry);
907 path.mnt = mntget(mnt);
908 error = prepare_open(path.dentry, oflag, ro, mode, name, attr);
910 struct file *file = dentry_open(&path, oflag, current_cred());
912 fd_install(fd, file);
914 error = PTR_ERR(file);
922 inode_unlock(d_inode(root));
930 SYSCALL_DEFINE4(mq_open, const char __user *, u_name, int, oflag, umode_t, mode,
931 struct mq_attr __user *, u_attr)
934 if (u_attr && copy_from_user(&attr, u_attr, sizeof(struct mq_attr)))
937 return do_mq_open(u_name, oflag, mode, u_attr ? &attr : NULL);
940 SYSCALL_DEFINE1(mq_unlink, const char __user *, u_name)
943 struct filename *name;
944 struct dentry *dentry;
945 struct inode *inode = NULL;
946 struct ipc_namespace *ipc_ns = current->nsproxy->ipc_ns;
947 struct vfsmount *mnt = ipc_ns->mq_mnt;
949 name = getname(u_name);
951 return PTR_ERR(name);
953 audit_inode_parent_hidden(name, mnt->mnt_root);
954 err = mnt_want_write(mnt);
957 inode_lock_nested(d_inode(mnt->mnt_root), I_MUTEX_PARENT);
958 dentry = lookup_one_len(name->name, mnt->mnt_root,
960 if (IS_ERR(dentry)) {
961 err = PTR_ERR(dentry);
965 inode = d_inode(dentry);
970 err = vfs_unlink(&init_user_ns, d_inode(dentry->d_parent),
976 inode_unlock(d_inode(mnt->mnt_root));
986 /* Pipelined send and receive functions.
988 * If a receiver finds no waiting message, then it registers itself in the
989 * list of waiting receivers. A sender checks that list before adding the new
990 * message into the message array. If there is a waiting receiver, then it
991 * bypasses the message array and directly hands the message over to the
992 * receiver. The receiver accepts the message and returns without grabbing the
995 * - Set pointer to message.
996 * - Queue the receiver task for later wakeup (without the info->lock).
997 * - Update its state to STATE_READY. Now the receiver can continue.
998 * - Wake up the process after the lock is dropped. Should the process wake up
999 * before this wakeup (due to a timeout or a signal) it will either see
1000 * STATE_READY and continue or acquire the lock to check the state again.
1002 * The same algorithm is used for senders.
1005 static inline void __pipelined_op(struct wake_q_head *wake_q,
1006 struct mqueue_inode_info *info,
1007 struct ext_wait_queue *this)
1009 struct task_struct *task;
1011 list_del(&this->list);
1012 task = get_task_struct(this->task);
1014 /* see MQ_BARRIER for purpose/pairing */
1015 smp_store_release(&this->state, STATE_READY);
1016 wake_q_add_safe(wake_q, task);
1019 /* pipelined_send() - send a message directly to the task waiting in
1020 * sys_mq_timedreceive() (without inserting message into a queue).
1022 static inline void pipelined_send(struct wake_q_head *wake_q,
1023 struct mqueue_inode_info *info,
1024 struct msg_msg *message,
1025 struct ext_wait_queue *receiver)
1027 receiver->msg = message;
1028 __pipelined_op(wake_q, info, receiver);
1031 /* pipelined_receive() - if there is task waiting in sys_mq_timedsend()
1032 * gets its message and put to the queue (we have one free place for sure). */
1033 static inline void pipelined_receive(struct wake_q_head *wake_q,
1034 struct mqueue_inode_info *info)
1036 struct ext_wait_queue *sender = wq_get_first_waiter(info, SEND);
1040 wake_up_interruptible(&info->wait_q);
1043 if (msg_insert(sender->msg, info))
1046 __pipelined_op(wake_q, info, sender);
1049 static int do_mq_timedsend(mqd_t mqdes, const char __user *u_msg_ptr,
1050 size_t msg_len, unsigned int msg_prio,
1051 struct timespec64 *ts)
1054 struct inode *inode;
1055 struct ext_wait_queue wait;
1056 struct ext_wait_queue *receiver;
1057 struct msg_msg *msg_ptr;
1058 struct mqueue_inode_info *info;
1059 ktime_t expires, *timeout = NULL;
1060 struct posix_msg_tree_node *new_leaf = NULL;
1062 DEFINE_WAKE_Q(wake_q);
1064 if (unlikely(msg_prio >= (unsigned long) MQ_PRIO_MAX))
1068 expires = timespec64_to_ktime(*ts);
1072 audit_mq_sendrecv(mqdes, msg_len, msg_prio, ts);
1075 if (unlikely(!f.file)) {
1080 inode = file_inode(f.file);
1081 if (unlikely(f.file->f_op != &mqueue_file_operations)) {
1085 info = MQUEUE_I(inode);
1088 if (unlikely(!(f.file->f_mode & FMODE_WRITE))) {
1093 if (unlikely(msg_len > info->attr.mq_msgsize)) {
1098 /* First try to allocate memory, before doing anything with
1099 * existing queues. */
1100 msg_ptr = load_msg(u_msg_ptr, msg_len);
1101 if (IS_ERR(msg_ptr)) {
1102 ret = PTR_ERR(msg_ptr);
1105 msg_ptr->m_ts = msg_len;
1106 msg_ptr->m_type = msg_prio;
1109 * msg_insert really wants us to have a valid, spare node struct so
1110 * it doesn't have to kmalloc a GFP_ATOMIC allocation, but it will
1111 * fall back to that if necessary.
1113 if (!info->node_cache)
1114 new_leaf = kmalloc(sizeof(*new_leaf), GFP_KERNEL);
1116 spin_lock(&info->lock);
1118 if (!info->node_cache && new_leaf) {
1119 /* Save our speculative allocation into the cache */
1120 INIT_LIST_HEAD(&new_leaf->msg_list);
1121 info->node_cache = new_leaf;
1127 if (info->attr.mq_curmsgs == info->attr.mq_maxmsg) {
1128 if (f.file->f_flags & O_NONBLOCK) {
1131 wait.task = current;
1132 wait.msg = (void *) msg_ptr;
1134 /* memory barrier not required, we hold info->lock */
1135 WRITE_ONCE(wait.state, STATE_NONE);
1136 ret = wq_sleep(info, SEND, timeout, &wait);
1138 * wq_sleep must be called with info->lock held, and
1139 * returns with the lock released
1144 receiver = wq_get_first_waiter(info, RECV);
1146 pipelined_send(&wake_q, info, msg_ptr, receiver);
1148 /* adds message to the queue */
1149 ret = msg_insert(msg_ptr, info);
1154 inode->i_atime = inode->i_mtime = inode->i_ctime =
1155 current_time(inode);
1158 spin_unlock(&info->lock);
1169 static int do_mq_timedreceive(mqd_t mqdes, char __user *u_msg_ptr,
1170 size_t msg_len, unsigned int __user *u_msg_prio,
1171 struct timespec64 *ts)
1174 struct msg_msg *msg_ptr;
1176 struct inode *inode;
1177 struct mqueue_inode_info *info;
1178 struct ext_wait_queue wait;
1179 ktime_t expires, *timeout = NULL;
1180 struct posix_msg_tree_node *new_leaf = NULL;
1183 expires = timespec64_to_ktime(*ts);
1187 audit_mq_sendrecv(mqdes, msg_len, 0, ts);
1190 if (unlikely(!f.file)) {
1195 inode = file_inode(f.file);
1196 if (unlikely(f.file->f_op != &mqueue_file_operations)) {
1200 info = MQUEUE_I(inode);
1203 if (unlikely(!(f.file->f_mode & FMODE_READ))) {
1208 /* checks if buffer is big enough */
1209 if (unlikely(msg_len < info->attr.mq_msgsize)) {
1215 * msg_insert really wants us to have a valid, spare node struct so
1216 * it doesn't have to kmalloc a GFP_ATOMIC allocation, but it will
1217 * fall back to that if necessary.
1219 if (!info->node_cache)
1220 new_leaf = kmalloc(sizeof(*new_leaf), GFP_KERNEL);
1222 spin_lock(&info->lock);
1224 if (!info->node_cache && new_leaf) {
1225 /* Save our speculative allocation into the cache */
1226 INIT_LIST_HEAD(&new_leaf->msg_list);
1227 info->node_cache = new_leaf;
1232 if (info->attr.mq_curmsgs == 0) {
1233 if (f.file->f_flags & O_NONBLOCK) {
1234 spin_unlock(&info->lock);
1237 wait.task = current;
1239 /* memory barrier not required, we hold info->lock */
1240 WRITE_ONCE(wait.state, STATE_NONE);
1241 ret = wq_sleep(info, RECV, timeout, &wait);
1245 DEFINE_WAKE_Q(wake_q);
1247 msg_ptr = msg_get(info);
1249 inode->i_atime = inode->i_mtime = inode->i_ctime =
1250 current_time(inode);
1252 /* There is now free space in queue. */
1253 pipelined_receive(&wake_q, info);
1254 spin_unlock(&info->lock);
1259 ret = msg_ptr->m_ts;
1261 if ((u_msg_prio && put_user(msg_ptr->m_type, u_msg_prio)) ||
1262 store_msg(u_msg_ptr, msg_ptr, msg_ptr->m_ts)) {
1273 SYSCALL_DEFINE5(mq_timedsend, mqd_t, mqdes, const char __user *, u_msg_ptr,
1274 size_t, msg_len, unsigned int, msg_prio,
1275 const struct __kernel_timespec __user *, u_abs_timeout)
1277 struct timespec64 ts, *p = NULL;
1278 if (u_abs_timeout) {
1279 int res = prepare_timeout(u_abs_timeout, &ts);
1284 return do_mq_timedsend(mqdes, u_msg_ptr, msg_len, msg_prio, p);
1287 SYSCALL_DEFINE5(mq_timedreceive, mqd_t, mqdes, char __user *, u_msg_ptr,
1288 size_t, msg_len, unsigned int __user *, u_msg_prio,
1289 const struct __kernel_timespec __user *, u_abs_timeout)
1291 struct timespec64 ts, *p = NULL;
1292 if (u_abs_timeout) {
1293 int res = prepare_timeout(u_abs_timeout, &ts);
1298 return do_mq_timedreceive(mqdes, u_msg_ptr, msg_len, u_msg_prio, p);
1302 * Notes: the case when user wants us to deregister (with NULL as pointer)
1303 * and he isn't currently owner of notification, will be silently discarded.
1304 * It isn't explicitly defined in the POSIX.
1306 static int do_mq_notify(mqd_t mqdes, const struct sigevent *notification)
1311 struct inode *inode;
1312 struct mqueue_inode_info *info;
1315 audit_mq_notify(mqdes, notification);
1319 if (notification != NULL) {
1320 if (unlikely(notification->sigev_notify != SIGEV_NONE &&
1321 notification->sigev_notify != SIGEV_SIGNAL &&
1322 notification->sigev_notify != SIGEV_THREAD))
1324 if (notification->sigev_notify == SIGEV_SIGNAL &&
1325 !valid_signal(notification->sigev_signo)) {
1328 if (notification->sigev_notify == SIGEV_THREAD) {
1331 /* create the notify skb */
1332 nc = alloc_skb(NOTIFY_COOKIE_LEN, GFP_KERNEL);
1336 if (copy_from_user(nc->data,
1337 notification->sigev_value.sival_ptr,
1338 NOTIFY_COOKIE_LEN)) {
1343 /* TODO: add a header? */
1344 skb_put(nc, NOTIFY_COOKIE_LEN);
1345 /* and attach it to the socket */
1347 f = fdget(notification->sigev_signo);
1352 sock = netlink_getsockbyfilp(f.file);
1355 ret = PTR_ERR(sock);
1359 timeo = MAX_SCHEDULE_TIMEOUT;
1360 ret = netlink_attachskb(sock, nc, &timeo, NULL);
1376 inode = file_inode(f.file);
1377 if (unlikely(f.file->f_op != &mqueue_file_operations)) {
1381 info = MQUEUE_I(inode);
1384 spin_lock(&info->lock);
1385 if (notification == NULL) {
1386 if (info->notify_owner == task_tgid(current)) {
1387 remove_notification(info);
1388 inode->i_atime = inode->i_ctime = current_time(inode);
1390 } else if (info->notify_owner != NULL) {
1393 switch (notification->sigev_notify) {
1395 info->notify.sigev_notify = SIGEV_NONE;
1398 info->notify_sock = sock;
1399 info->notify_cookie = nc;
1402 info->notify.sigev_notify = SIGEV_THREAD;
1405 info->notify.sigev_signo = notification->sigev_signo;
1406 info->notify.sigev_value = notification->sigev_value;
1407 info->notify.sigev_notify = SIGEV_SIGNAL;
1408 info->notify_self_exec_id = current->self_exec_id;
1412 info->notify_owner = get_pid(task_tgid(current));
1413 info->notify_user_ns = get_user_ns(current_user_ns());
1414 inode->i_atime = inode->i_ctime = current_time(inode);
1416 spin_unlock(&info->lock);
1421 netlink_detachskb(sock, nc);
1429 SYSCALL_DEFINE2(mq_notify, mqd_t, mqdes,
1430 const struct sigevent __user *, u_notification)
1432 struct sigevent n, *p = NULL;
1433 if (u_notification) {
1434 if (copy_from_user(&n, u_notification, sizeof(struct sigevent)))
1438 return do_mq_notify(mqdes, p);
1441 static int do_mq_getsetattr(int mqdes, struct mq_attr *new, struct mq_attr *old)
1444 struct inode *inode;
1445 struct mqueue_inode_info *info;
1447 if (new && (new->mq_flags & (~O_NONBLOCK)))
1454 if (unlikely(f.file->f_op != &mqueue_file_operations)) {
1459 inode = file_inode(f.file);
1460 info = MQUEUE_I(inode);
1462 spin_lock(&info->lock);
1466 old->mq_flags = f.file->f_flags & O_NONBLOCK;
1469 audit_mq_getsetattr(mqdes, new);
1470 spin_lock(&f.file->f_lock);
1471 if (new->mq_flags & O_NONBLOCK)
1472 f.file->f_flags |= O_NONBLOCK;
1474 f.file->f_flags &= ~O_NONBLOCK;
1475 spin_unlock(&f.file->f_lock);
1477 inode->i_atime = inode->i_ctime = current_time(inode);
1480 spin_unlock(&info->lock);
1485 SYSCALL_DEFINE3(mq_getsetattr, mqd_t, mqdes,
1486 const struct mq_attr __user *, u_mqstat,
1487 struct mq_attr __user *, u_omqstat)
1490 struct mq_attr mqstat, omqstat;
1491 struct mq_attr *new = NULL, *old = NULL;
1495 if (copy_from_user(new, u_mqstat, sizeof(struct mq_attr)))
1501 ret = do_mq_getsetattr(mqdes, new, old);
1505 if (copy_to_user(u_omqstat, old, sizeof(struct mq_attr)))
1510 #ifdef CONFIG_COMPAT
1512 struct compat_mq_attr {
1513 compat_long_t mq_flags; /* message queue flags */
1514 compat_long_t mq_maxmsg; /* maximum number of messages */
1515 compat_long_t mq_msgsize; /* maximum message size */
1516 compat_long_t mq_curmsgs; /* number of messages currently queued */
1517 compat_long_t __reserved[4]; /* ignored for input, zeroed for output */
1520 static inline int get_compat_mq_attr(struct mq_attr *attr,
1521 const struct compat_mq_attr __user *uattr)
1523 struct compat_mq_attr v;
1525 if (copy_from_user(&v, uattr, sizeof(*uattr)))
1528 memset(attr, 0, sizeof(*attr));
1529 attr->mq_flags = v.mq_flags;
1530 attr->mq_maxmsg = v.mq_maxmsg;
1531 attr->mq_msgsize = v.mq_msgsize;
1532 attr->mq_curmsgs = v.mq_curmsgs;
1536 static inline int put_compat_mq_attr(const struct mq_attr *attr,
1537 struct compat_mq_attr __user *uattr)
1539 struct compat_mq_attr v;
1541 memset(&v, 0, sizeof(v));
1542 v.mq_flags = attr->mq_flags;
1543 v.mq_maxmsg = attr->mq_maxmsg;
1544 v.mq_msgsize = attr->mq_msgsize;
1545 v.mq_curmsgs = attr->mq_curmsgs;
1546 if (copy_to_user(uattr, &v, sizeof(*uattr)))
1551 COMPAT_SYSCALL_DEFINE4(mq_open, const char __user *, u_name,
1552 int, oflag, compat_mode_t, mode,
1553 struct compat_mq_attr __user *, u_attr)
1555 struct mq_attr attr, *p = NULL;
1556 if (u_attr && oflag & O_CREAT) {
1558 if (get_compat_mq_attr(&attr, u_attr))
1561 return do_mq_open(u_name, oflag, mode, p);
1564 COMPAT_SYSCALL_DEFINE2(mq_notify, mqd_t, mqdes,
1565 const struct compat_sigevent __user *, u_notification)
1567 struct sigevent n, *p = NULL;
1568 if (u_notification) {
1569 if (get_compat_sigevent(&n, u_notification))
1571 if (n.sigev_notify == SIGEV_THREAD)
1572 n.sigev_value.sival_ptr = compat_ptr(n.sigev_value.sival_int);
1575 return do_mq_notify(mqdes, p);
1578 COMPAT_SYSCALL_DEFINE3(mq_getsetattr, mqd_t, mqdes,
1579 const struct compat_mq_attr __user *, u_mqstat,
1580 struct compat_mq_attr __user *, u_omqstat)
1583 struct mq_attr mqstat, omqstat;
1584 struct mq_attr *new = NULL, *old = NULL;
1588 if (get_compat_mq_attr(new, u_mqstat))
1594 ret = do_mq_getsetattr(mqdes, new, old);
1598 if (put_compat_mq_attr(old, u_omqstat))
1604 #ifdef CONFIG_COMPAT_32BIT_TIME
1605 static int compat_prepare_timeout(const struct old_timespec32 __user *p,
1606 struct timespec64 *ts)
1608 if (get_old_timespec32(ts, p))
1610 if (!timespec64_valid(ts))
1615 SYSCALL_DEFINE5(mq_timedsend_time32, mqd_t, mqdes,
1616 const char __user *, u_msg_ptr,
1617 unsigned int, msg_len, unsigned int, msg_prio,
1618 const struct old_timespec32 __user *, u_abs_timeout)
1620 struct timespec64 ts, *p = NULL;
1621 if (u_abs_timeout) {
1622 int res = compat_prepare_timeout(u_abs_timeout, &ts);
1627 return do_mq_timedsend(mqdes, u_msg_ptr, msg_len, msg_prio, p);
1630 SYSCALL_DEFINE5(mq_timedreceive_time32, mqd_t, mqdes,
1631 char __user *, u_msg_ptr,
1632 unsigned int, msg_len, unsigned int __user *, u_msg_prio,
1633 const struct old_timespec32 __user *, u_abs_timeout)
1635 struct timespec64 ts, *p = NULL;
1636 if (u_abs_timeout) {
1637 int res = compat_prepare_timeout(u_abs_timeout, &ts);
1642 return do_mq_timedreceive(mqdes, u_msg_ptr, msg_len, u_msg_prio, p);
1646 static const struct inode_operations mqueue_dir_inode_operations = {
1647 .lookup = simple_lookup,
1648 .create = mqueue_create,
1649 .unlink = mqueue_unlink,
1652 static const struct file_operations mqueue_file_operations = {
1653 .flush = mqueue_flush_file,
1654 .poll = mqueue_poll_file,
1655 .read = mqueue_read_file,
1656 .llseek = default_llseek,
1659 static const struct super_operations mqueue_super_ops = {
1660 .alloc_inode = mqueue_alloc_inode,
1661 .free_inode = mqueue_free_inode,
1662 .evict_inode = mqueue_evict_inode,
1663 .statfs = simple_statfs,
1666 static const struct fs_context_operations mqueue_fs_context_ops = {
1667 .free = mqueue_fs_context_free,
1668 .get_tree = mqueue_get_tree,
1671 static struct file_system_type mqueue_fs_type = {
1673 .init_fs_context = mqueue_init_fs_context,
1674 .kill_sb = kill_litter_super,
1675 .fs_flags = FS_USERNS_MOUNT,
1678 int mq_init_ns(struct ipc_namespace *ns)
1682 ns->mq_queues_count = 0;
1683 ns->mq_queues_max = DFLT_QUEUESMAX;
1684 ns->mq_msg_max = DFLT_MSGMAX;
1685 ns->mq_msgsize_max = DFLT_MSGSIZEMAX;
1686 ns->mq_msg_default = DFLT_MSG;
1687 ns->mq_msgsize_default = DFLT_MSGSIZE;
1689 m = mq_create_mount(ns);
1696 void mq_clear_sbinfo(struct ipc_namespace *ns)
1698 ns->mq_mnt->mnt_sb->s_fs_info = NULL;
1701 void mq_put_mnt(struct ipc_namespace *ns)
1703 kern_unmount(ns->mq_mnt);
1706 static int __init init_mqueue_fs(void)
1710 mqueue_inode_cachep = kmem_cache_create("mqueue_inode_cache",
1711 sizeof(struct mqueue_inode_info), 0,
1712 SLAB_HWCACHE_ALIGN|SLAB_ACCOUNT, init_once);
1713 if (mqueue_inode_cachep == NULL)
1716 /* ignore failures - they are not fatal */
1717 mq_sysctl_table = mq_register_sysctl_table();
1719 error = register_filesystem(&mqueue_fs_type);
1723 spin_lock_init(&mq_lock);
1725 error = mq_init_ns(&init_ipc_ns);
1727 goto out_filesystem;
1732 unregister_filesystem(&mqueue_fs_type);
1734 if (mq_sysctl_table)
1735 unregister_sysctl_table(mq_sysctl_table);
1736 kmem_cache_destroy(mqueue_inode_cachep);
1740 device_initcall(init_mqueue_fs);