blob: 28bd64ddeda3f438742d5297e13da0269ad62609 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * POSIX message queues filesystem for Linux.
3 *
4 * Copyright (C) 2003,2004 Krzysztof Benedyczak (golbi@mat.uni.torun.pl)
Michal Wronskif66e9282006-10-03 23:23:27 +02005 * Michal Wronski (michal.wronski@gmail.com)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006 *
7 * Spinlocks: Mohamed Abbas (abbas.mohamed@intel.com)
8 * Lockless receive & send, fd based notify:
9 * Manfred Spraul (manfred@colorfullife.com)
10 *
George C. Wilson20ca73b2006-05-24 16:09:55 -050011 * Audit: George Wilson (ltcgcw@us.ibm.com)
12 *
Linus Torvalds1da177e2005-04-16 15:20:36 -070013 * This file is released under the GPL.
14 */
15
Randy.Dunlapc59ede72006-01-11 12:17:46 -080016#include <linux/capability.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070017#include <linux/init.h>
18#include <linux/pagemap.h>
19#include <linux/file.h>
20#include <linux/mount.h>
21#include <linux/namei.h>
22#include <linux/sysctl.h>
23#include <linux/poll.h>
24#include <linux/mqueue.h>
25#include <linux/msg.h>
26#include <linux/skbuff.h>
27#include <linux/netlink.h>
28#include <linux/syscalls.h>
George C. Wilson20ca73b2006-05-24 16:09:55 -050029#include <linux/audit.h>
Jesper Juhl7ed20e12005-05-01 08:59:14 -070030#include <linux/signal.h>
Ingo Molnar5f921ae2006-03-26 01:37:17 -080031#include <linux/mutex.h>
Pavel Emelyanovb4888932007-10-18 23:40:14 -070032#include <linux/nsproxy.h>
33#include <linux/pid.h>
Serge E. Hallyn614b84c2009-04-06 19:01:08 -070034#include <linux/ipc_namespace.h>
Serge E. Hallyn6b550f92012-01-10 15:11:37 -080035#include <linux/user_namespace.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090036#include <linux/slab.h>
Ingo Molnar5f921ae2006-03-26 01:37:17 -080037
Linus Torvalds1da177e2005-04-16 15:20:36 -070038#include <net/sock.h>
39#include "util.h"
40
41#define MQUEUE_MAGIC 0x19800202
42#define DIRENT_SIZE 20
43#define FILENT_SIZE 80
44
45#define SEND 0
46#define RECV 1
47
48#define STATE_NONE 0
49#define STATE_PENDING 1
50#define STATE_READY 2
51
Linus Torvalds1da177e2005-04-16 15:20:36 -070052struct ext_wait_queue { /* queue of sleeping tasks */
53 struct task_struct *task;
54 struct list_head list;
55 struct msg_msg *msg; /* ptr of loaded message */
56 int state; /* one of STATE_* values */
57};
58
59struct mqueue_inode_info {
60 spinlock_t lock;
61 struct inode vfs_inode;
62 wait_queue_head_t wait_q;
63
64 struct msg_msg **messages;
65 struct mq_attr attr;
66
67 struct sigevent notify;
Cedric Le Goatera03fcb72006-10-02 02:17:26 -070068 struct pid* notify_owner;
Adrian Bunk338cec32005-09-10 00:26:54 -070069 struct user_struct *user; /* user who created, for accounting */
Linus Torvalds1da177e2005-04-16 15:20:36 -070070 struct sock *notify_sock;
71 struct sk_buff *notify_cookie;
72
73 /* for tasks waiting for free space and messages, respectively */
74 struct ext_wait_queue e_wait_q[2];
75
76 unsigned long qsize; /* size of queue in memory (sum of all msgs) */
77};
78
Arjan van de Ven92e1d5b2007-02-12 00:55:39 -080079static const struct inode_operations mqueue_dir_inode_operations;
Arjan van de Ven9a321442007-02-12 00:55:35 -080080static const struct file_operations mqueue_file_operations;
Alexey Dobriyanb87221d2009-09-21 17:01:09 -070081static const struct super_operations mqueue_super_ops;
Linus Torvalds1da177e2005-04-16 15:20:36 -070082static void remove_notification(struct mqueue_inode_info *info);
83
Christoph Lametere18b8902006-12-06 20:33:20 -080084static struct kmem_cache *mqueue_inode_cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -070085
86static struct ctl_table_header * mq_sysctl_table;
87
88static inline struct mqueue_inode_info *MQUEUE_I(struct inode *inode)
89{
90 return container_of(inode, struct mqueue_inode_info, vfs_inode);
91}
92
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -070093/*
94 * This routine should be called with the mq_lock held.
95 */
96static inline struct ipc_namespace *__get_ns_from_inode(struct inode *inode)
Serge E. Hallyn614b84c2009-04-06 19:01:08 -070097{
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -070098 return get_ipc_ns(inode->i_sb->s_fs_info);
Serge E. Hallyn614b84c2009-04-06 19:01:08 -070099}
100
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700101static struct ipc_namespace *get_ns_from_inode(struct inode *inode)
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700102{
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700103 struct ipc_namespace *ns;
104
105 spin_lock(&mq_lock);
106 ns = __get_ns_from_inode(inode);
107 spin_unlock(&mq_lock);
108 return ns;
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700109}
110
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700111static struct inode *mqueue_get_inode(struct super_block *sb,
Al Viro1b9d5ff2011-07-24 14:18:20 -0400112 struct ipc_namespace *ipc_ns, umode_t mode,
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700113 struct mq_attr *attr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114{
David Howells86a264a2008-11-14 10:39:18 +1100115 struct user_struct *u = current_user();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116 struct inode *inode;
Jiri Slabyd40dcdb2011-07-26 16:08:47 -0700117 int ret = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700118
119 inode = new_inode(sb);
Jiri Slaby04715202011-07-26 16:08:46 -0700120 if (!inode)
121 goto err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122
Jiri Slaby04715202011-07-26 16:08:46 -0700123 inode->i_ino = get_next_ino();
124 inode->i_mode = mode;
125 inode->i_uid = current_fsuid();
126 inode->i_gid = current_fsgid();
127 inode->i_mtime = inode->i_ctime = inode->i_atime = CURRENT_TIME;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700128
Jiri Slaby04715202011-07-26 16:08:46 -0700129 if (S_ISREG(mode)) {
130 struct mqueue_inode_info *info;
Jiri Slaby04715202011-07-26 16:08:46 -0700131 unsigned long mq_bytes, mq_msg_tblsz;
André Goddard Rosac8308b12010-02-23 04:04:23 -0300132
Jiri Slaby04715202011-07-26 16:08:46 -0700133 inode->i_fop = &mqueue_file_operations;
134 inode->i_size = FILENT_SIZE;
135 /* mqueue specific info */
136 info = MQUEUE_I(inode);
137 spin_lock_init(&info->lock);
138 init_waitqueue_head(&info->wait_q);
139 INIT_LIST_HEAD(&info->e_wait_q[0].list);
140 INIT_LIST_HEAD(&info->e_wait_q[1].list);
141 info->notify_owner = NULL;
142 info->qsize = 0;
143 info->user = NULL; /* set when all is ok */
144 memset(&info->attr, 0, sizeof(info->attr));
145 info->attr.mq_maxmsg = ipc_ns->mq_msg_max;
146 info->attr.mq_msgsize = ipc_ns->mq_msgsize_max;
147 if (attr) {
148 info->attr.mq_maxmsg = attr->mq_maxmsg;
149 info->attr.mq_msgsize = attr->mq_msgsize;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700150 }
Jiri Slaby04715202011-07-26 16:08:46 -0700151 mq_msg_tblsz = info->attr.mq_maxmsg * sizeof(struct msg_msg *);
152 info->messages = kmalloc(mq_msg_tblsz, GFP_KERNEL);
153 if (!info->messages)
154 goto out_inode;
155
156 mq_bytes = (mq_msg_tblsz +
157 (info->attr.mq_maxmsg * info->attr.mq_msgsize));
158
159 spin_lock(&mq_lock);
160 if (u->mq_bytes + mq_bytes < u->mq_bytes ||
Davidlohr Bueso2a4e64b2012-01-20 14:34:01 -0800161 u->mq_bytes + mq_bytes > rlimit(RLIMIT_MSGQUEUE)) {
Jiri Slaby04715202011-07-26 16:08:46 -0700162 spin_unlock(&mq_lock);
163 /* mqueue_evict_inode() releases info->messages */
Jiri Slabyd40dcdb2011-07-26 16:08:47 -0700164 ret = -EMFILE;
Jiri Slaby04715202011-07-26 16:08:46 -0700165 goto out_inode;
166 }
167 u->mq_bytes += mq_bytes;
168 spin_unlock(&mq_lock);
169
170 /* all is ok */
171 info->user = get_uid(u);
172 } else if (S_ISDIR(mode)) {
173 inc_nlink(inode);
174 /* Some things misbehave if size == 0 on a directory */
175 inode->i_size = 2 * DIRENT_SIZE;
176 inode->i_op = &mqueue_dir_inode_operations;
177 inode->i_fop = &simple_dir_operations;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178 }
Jiri Slaby04715202011-07-26 16:08:46 -0700179
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180 return inode;
181out_inode:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182 iput(inode);
Jiri Slaby04715202011-07-26 16:08:46 -0700183err:
Jiri Slabyd40dcdb2011-07-26 16:08:47 -0700184 return ERR_PTR(ret);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700185}
186
187static int mqueue_fill_super(struct super_block *sb, void *data, int silent)
188{
189 struct inode *inode;
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700190 struct ipc_namespace *ns = data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191
192 sb->s_blocksize = PAGE_CACHE_SIZE;
193 sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
194 sb->s_magic = MQUEUE_MAGIC;
195 sb->s_op = &mqueue_super_ops;
196
Al Viro48fde702012-01-08 22:15:13 -0500197 inode = mqueue_get_inode(sb, ns, S_IFDIR | S_ISVTX | S_IRWXUGO, NULL);
198 if (IS_ERR(inode))
199 return PTR_ERR(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700200
Al Viro48fde702012-01-08 22:15:13 -0500201 sb->s_root = d_make_root(inode);
202 if (!sb->s_root)
203 return -ENOMEM;
204 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205}
206
Al Viroceefda62010-07-26 13:16:50 +0400207static struct dentry *mqueue_mount(struct file_system_type *fs_type,
David Howells454e2392006-06-23 02:02:57 -0700208 int flags, const char *dev_name,
Al Viroceefda62010-07-26 13:16:50 +0400209 void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700210{
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700211 if (!(flags & MS_KERNMOUNT))
212 data = current->nsproxy->ipc_ns;
Al Viroceefda62010-07-26 13:16:50 +0400213 return mount_ns(fs_type, flags, data, mqueue_fill_super);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214}
215
Alexey Dobriyan51cc5062008-07-25 19:45:34 -0700216static void init_once(void *foo)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700217{
218 struct mqueue_inode_info *p = (struct mqueue_inode_info *) foo;
219
Christoph Lametera35afb82007-05-16 22:10:57 -0700220 inode_init_once(&p->vfs_inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700221}
222
223static struct inode *mqueue_alloc_inode(struct super_block *sb)
224{
225 struct mqueue_inode_info *ei;
226
Christoph Lametere94b1762006-12-06 20:33:17 -0800227 ei = kmem_cache_alloc(mqueue_inode_cachep, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228 if (!ei)
229 return NULL;
230 return &ei->vfs_inode;
231}
232
Nick Pigginfa0d7e32011-01-07 17:49:49 +1100233static void mqueue_i_callback(struct rcu_head *head)
234{
235 struct inode *inode = container_of(head, struct inode, i_rcu);
Nick Pigginfa0d7e32011-01-07 17:49:49 +1100236 kmem_cache_free(mqueue_inode_cachep, MQUEUE_I(inode));
237}
238
Linus Torvalds1da177e2005-04-16 15:20:36 -0700239static void mqueue_destroy_inode(struct inode *inode)
240{
Nick Pigginfa0d7e32011-01-07 17:49:49 +1100241 call_rcu(&inode->i_rcu, mqueue_i_callback);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242}
243
Al Viro6d8af642010-06-05 16:29:45 -0400244static void mqueue_evict_inode(struct inode *inode)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700245{
246 struct mqueue_inode_info *info;
247 struct user_struct *user;
248 unsigned long mq_bytes;
249 int i;
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700250 struct ipc_namespace *ipc_ns;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251
Al Viro6d8af642010-06-05 16:29:45 -0400252 end_writeback(inode);
253
254 if (S_ISDIR(inode->i_mode))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700255 return;
Al Viro6d8af642010-06-05 16:29:45 -0400256
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700257 ipc_ns = get_ns_from_inode(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700258 info = MQUEUE_I(inode);
259 spin_lock(&info->lock);
260 for (i = 0; i < info->attr.mq_curmsgs; i++)
261 free_msg(info->messages[i]);
262 kfree(info->messages);
263 spin_unlock(&info->lock);
264
André Goddard Rosa8834cf72010-02-23 04:04:24 -0300265 /* Total amount of bytes accounted for the mqueue */
266 mq_bytes = info->attr.mq_maxmsg * (sizeof(struct msg_msg *)
267 + info->attr.mq_msgsize);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268 user = info->user;
269 if (user) {
270 spin_lock(&mq_lock);
271 user->mq_bytes -= mq_bytes;
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700272 /*
273 * get_ns_from_inode() ensures that the
274 * (ipc_ns = sb->s_fs_info) is either a valid ipc_ns
275 * to which we now hold a reference, or it is NULL.
276 * We can't put it here under mq_lock, though.
277 */
278 if (ipc_ns)
279 ipc_ns->mq_queues_count--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280 spin_unlock(&mq_lock);
281 free_uid(user);
282 }
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700283 if (ipc_ns)
284 put_ipc_ns(ipc_ns);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285}
286
287static int mqueue_create(struct inode *dir, struct dentry *dentry,
Al Viro4acdaf22011-07-26 01:42:34 -0400288 umode_t mode, struct nameidata *nd)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700289{
290 struct inode *inode;
291 struct mq_attr *attr = dentry->d_fsdata;
292 int error;
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700293 struct ipc_namespace *ipc_ns;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294
295 spin_lock(&mq_lock);
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700296 ipc_ns = __get_ns_from_inode(dir);
297 if (!ipc_ns) {
298 error = -EACCES;
299 goto out_unlock;
300 }
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700301 if (ipc_ns->mq_queues_count >= ipc_ns->mq_queues_max &&
302 !capable(CAP_SYS_RESOURCE)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303 error = -ENOSPC;
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700304 goto out_unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700305 }
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700306 ipc_ns->mq_queues_count++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700307 spin_unlock(&mq_lock);
308
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700309 inode = mqueue_get_inode(dir->i_sb, ipc_ns, mode, attr);
Jiri Slabyd40dcdb2011-07-26 16:08:47 -0700310 if (IS_ERR(inode)) {
311 error = PTR_ERR(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700312 spin_lock(&mq_lock);
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700313 ipc_ns->mq_queues_count--;
314 goto out_unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700315 }
316
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700317 put_ipc_ns(ipc_ns);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318 dir->i_size += DIRENT_SIZE;
319 dir->i_ctime = dir->i_mtime = dir->i_atime = CURRENT_TIME;
320
321 d_instantiate(dentry, inode);
322 dget(dentry);
323 return 0;
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700324out_unlock:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325 spin_unlock(&mq_lock);
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700326 if (ipc_ns)
327 put_ipc_ns(ipc_ns);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328 return error;
329}
330
331static int mqueue_unlink(struct inode *dir, struct dentry *dentry)
332{
333 struct inode *inode = dentry->d_inode;
334
335 dir->i_ctime = dir->i_mtime = dir->i_atime = CURRENT_TIME;
336 dir->i_size -= DIRENT_SIZE;
Dave Hansen9a53c3a2006-09-30 23:29:03 -0700337 drop_nlink(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700338 dput(dentry);
339 return 0;
340}
341
342/*
343* This is routine for system read from queue file.
344* To avoid mess with doing here some sort of mq_receive we allow
345* to read only queue size & notification info (the only values
346* that are interesting from user point of view and aren't accessible
347* through std routines)
348*/
349static ssize_t mqueue_read_file(struct file *filp, char __user *u_data,
Akinobu Mitaf1a43f92008-07-25 01:48:07 -0700350 size_t count, loff_t *off)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700351{
Josef Sipek6d630792006-12-08 02:37:11 -0800352 struct mqueue_inode_info *info = MQUEUE_I(filp->f_path.dentry->d_inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353 char buffer[FILENT_SIZE];
Akinobu Mitaf1a43f92008-07-25 01:48:07 -0700354 ssize_t ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700355
356 spin_lock(&info->lock);
357 snprintf(buffer, sizeof(buffer),
358 "QSIZE:%-10lu NOTIFY:%-5d SIGNO:%-5d NOTIFY_PID:%-6d\n",
359 info->qsize,
360 info->notify_owner ? info->notify.sigev_notify : 0,
361 (info->notify_owner &&
362 info->notify.sigev_notify == SIGEV_SIGNAL) ?
363 info->notify.sigev_signo : 0,
Pavel Emelyanov6c5f3e72008-02-08 04:19:20 -0800364 pid_vnr(info->notify_owner));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700365 spin_unlock(&info->lock);
366 buffer[sizeof(buffer)-1] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367
Akinobu Mitaf1a43f92008-07-25 01:48:07 -0700368 ret = simple_read_from_buffer(u_data, count, off, buffer,
369 strlen(buffer));
370 if (ret <= 0)
371 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372
Josef Sipek6d630792006-12-08 02:37:11 -0800373 filp->f_path.dentry->d_inode->i_atime = filp->f_path.dentry->d_inode->i_ctime = CURRENT_TIME;
Akinobu Mitaf1a43f92008-07-25 01:48:07 -0700374 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375}
376
Miklos Szeredi75e1fcc2006-06-23 02:05:12 -0700377static int mqueue_flush_file(struct file *filp, fl_owner_t id)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378{
Josef Sipek6d630792006-12-08 02:37:11 -0800379 struct mqueue_inode_info *info = MQUEUE_I(filp->f_path.dentry->d_inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380
381 spin_lock(&info->lock);
Cedric Le Goatera03fcb72006-10-02 02:17:26 -0700382 if (task_tgid(current) == info->notify_owner)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700383 remove_notification(info);
384
385 spin_unlock(&info->lock);
386 return 0;
387}
388
389static unsigned int mqueue_poll_file(struct file *filp, struct poll_table_struct *poll_tab)
390{
Josef Sipek6d630792006-12-08 02:37:11 -0800391 struct mqueue_inode_info *info = MQUEUE_I(filp->f_path.dentry->d_inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392 int retval = 0;
393
394 poll_wait(filp, &info->wait_q, poll_tab);
395
396 spin_lock(&info->lock);
397 if (info->attr.mq_curmsgs)
398 retval = POLLIN | POLLRDNORM;
399
400 if (info->attr.mq_curmsgs < info->attr.mq_maxmsg)
401 retval |= POLLOUT | POLLWRNORM;
402 spin_unlock(&info->lock);
403
404 return retval;
405}
406
407/* Adds current to info->e_wait_q[sr] before element with smaller prio */
408static void wq_add(struct mqueue_inode_info *info, int sr,
409 struct ext_wait_queue *ewp)
410{
411 struct ext_wait_queue *walk;
412
413 ewp->task = current;
414
415 list_for_each_entry(walk, &info->e_wait_q[sr].list, list) {
416 if (walk->task->static_prio <= current->static_prio) {
417 list_add_tail(&ewp->list, &walk->list);
418 return;
419 }
420 }
421 list_add_tail(&ewp->list, &info->e_wait_q[sr].list);
422}
423
424/*
425 * Puts current task to sleep. Caller must hold queue lock. After return
426 * lock isn't held.
427 * sr: SEND or RECV
428 */
429static int wq_sleep(struct mqueue_inode_info *info, int sr,
Carsten Emde9ca7d8e2010-04-02 22:40:20 +0200430 ktime_t *timeout, struct ext_wait_queue *ewp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431{
432 int retval;
433 signed long time;
434
435 wq_add(info, sr, ewp);
436
437 for (;;) {
438 set_current_state(TASK_INTERRUPTIBLE);
439
440 spin_unlock(&info->lock);
Wanlong Gao32ea8452011-10-31 17:06:35 -0700441 time = schedule_hrtimeout_range_clock(timeout, 0,
442 HRTIMER_MODE_ABS, CLOCK_REALTIME);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443
444 while (ewp->state == STATE_PENDING)
445 cpu_relax();
446
447 if (ewp->state == STATE_READY) {
448 retval = 0;
449 goto out;
450 }
451 spin_lock(&info->lock);
452 if (ewp->state == STATE_READY) {
453 retval = 0;
454 goto out_unlock;
455 }
456 if (signal_pending(current)) {
457 retval = -ERESTARTSYS;
458 break;
459 }
460 if (time == 0) {
461 retval = -ETIMEDOUT;
462 break;
463 }
464 }
465 list_del(&ewp->list);
466out_unlock:
467 spin_unlock(&info->lock);
468out:
469 return retval;
470}
471
472/*
473 * Returns waiting task that should be serviced first or NULL if none exists
474 */
475static struct ext_wait_queue *wq_get_first_waiter(
476 struct mqueue_inode_info *info, int sr)
477{
478 struct list_head *ptr;
479
480 ptr = info->e_wait_q[sr].list.prev;
481 if (ptr == &info->e_wait_q[sr].list)
482 return NULL;
483 return list_entry(ptr, struct ext_wait_queue, list);
484}
485
486/* Auxiliary functions to manipulate messages' list */
487static void msg_insert(struct msg_msg *ptr, struct mqueue_inode_info *info)
488{
489 int k;
490
491 k = info->attr.mq_curmsgs - 1;
492 while (k >= 0 && info->messages[k]->m_type >= ptr->m_type) {
493 info->messages[k + 1] = info->messages[k];
494 k--;
495 }
496 info->attr.mq_curmsgs++;
497 info->qsize += ptr->m_ts;
498 info->messages[k + 1] = ptr;
499}
500
501static inline struct msg_msg *msg_get(struct mqueue_inode_info *info)
502{
503 info->qsize -= info->messages[--info->attr.mq_curmsgs]->m_ts;
504 return info->messages[info->attr.mq_curmsgs];
505}
506
507static inline void set_cookie(struct sk_buff *skb, char code)
508{
509 ((char*)skb->data)[NOTIFY_COOKIE_LEN-1] = code;
510}
511
512/*
513 * The next function is only to split too long sys_mq_timedsend
514 */
515static void __do_notify(struct mqueue_inode_info *info)
516{
517 /* notification
518 * invoked when there is registered process and there isn't process
519 * waiting synchronously for message AND state of queue changed from
520 * empty to not empty. Here we are sure that no one is waiting
521 * synchronously. */
522 if (info->notify_owner &&
523 info->attr.mq_curmsgs == 1) {
524 struct siginfo sig_i;
525 switch (info->notify.sigev_notify) {
526 case SIGEV_NONE:
527 break;
528 case SIGEV_SIGNAL:
529 /* sends signal */
530
531 sig_i.si_signo = info->notify.sigev_signo;
532 sig_i.si_errno = 0;
533 sig_i.si_code = SI_MESGQ;
534 sig_i.si_value = info->notify.sigev_value;
Serge E. Hallyn6b550f92012-01-10 15:11:37 -0800535 /* map current pid/uid into info->owner's namespaces */
536 rcu_read_lock();
Sukadev Bhattiprolua6684992009-01-07 18:08:50 -0800537 sig_i.si_pid = task_tgid_nr_ns(current,
538 ns_of_pid(info->notify_owner));
Serge E. Hallyn6b550f92012-01-10 15:11:37 -0800539 sig_i.si_uid = user_ns_map_uid(info->user->user_ns,
540 current_cred(), current_uid());
541 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542
Cedric Le Goatera03fcb72006-10-02 02:17:26 -0700543 kill_pid_info(info->notify.sigev_signo,
544 &sig_i, info->notify_owner);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545 break;
546 case SIGEV_THREAD:
547 set_cookie(info->notify_cookie, NOTIFY_WOKENUP);
Denis V. Lunev7ee015e2007-10-10 21:14:03 -0700548 netlink_sendskb(info->notify_sock, info->notify_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549 break;
550 }
551 /* after notification unregisters process */
Cedric Le Goatera03fcb72006-10-02 02:17:26 -0700552 put_pid(info->notify_owner);
553 info->notify_owner = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700554 }
555 wake_up(&info->wait_q);
556}
557
Carsten Emde9ca7d8e2010-04-02 22:40:20 +0200558static int prepare_timeout(const struct timespec __user *u_abs_timeout,
559 ktime_t *expires, struct timespec *ts)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700560{
Carsten Emde9ca7d8e2010-04-02 22:40:20 +0200561 if (copy_from_user(ts, u_abs_timeout, sizeof(struct timespec)))
562 return -EFAULT;
563 if (!timespec_valid(ts))
564 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565
Carsten Emde9ca7d8e2010-04-02 22:40:20 +0200566 *expires = timespec_to_ktime(*ts);
567 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568}
569
570static void remove_notification(struct mqueue_inode_info *info)
571{
Cedric Le Goatera03fcb72006-10-02 02:17:26 -0700572 if (info->notify_owner != NULL &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573 info->notify.sigev_notify == SIGEV_THREAD) {
574 set_cookie(info->notify_cookie, NOTIFY_REMOVED);
Denis V. Lunev7ee015e2007-10-10 21:14:03 -0700575 netlink_sendskb(info->notify_sock, info->notify_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700576 }
Cedric Le Goatera03fcb72006-10-02 02:17:26 -0700577 put_pid(info->notify_owner);
578 info->notify_owner = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579}
580
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700581static int mq_attr_ok(struct ipc_namespace *ipc_ns, struct mq_attr *attr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700582{
583 if (attr->mq_maxmsg <= 0 || attr->mq_msgsize <= 0)
584 return 0;
585 if (capable(CAP_SYS_RESOURCE)) {
586 if (attr->mq_maxmsg > HARD_MSGMAX)
587 return 0;
588 } else {
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700589 if (attr->mq_maxmsg > ipc_ns->mq_msg_max ||
590 attr->mq_msgsize > ipc_ns->mq_msgsize_max)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700591 return 0;
592 }
593 /* check for overflow */
594 if (attr->mq_msgsize > ULONG_MAX/attr->mq_maxmsg)
595 return 0;
André Goddard Rosa8834cf72010-02-23 04:04:24 -0300596 if ((unsigned long)(attr->mq_maxmsg * (attr->mq_msgsize
597 + sizeof (struct msg_msg *))) <
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598 (unsigned long)(attr->mq_maxmsg * attr->mq_msgsize))
599 return 0;
600 return 1;
601}
602
603/*
604 * Invoked when creating a new queue via sys_mq_open
605 */
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700606static struct file *do_create(struct ipc_namespace *ipc_ns, struct dentry *dir,
Al Viro4acdaf22011-07-26 01:42:34 -0400607 struct dentry *dentry, int oflag, umode_t mode,
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700608 struct mq_attr *attr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609{
David Howells745ca242008-11-14 10:39:22 +1100610 const struct cred *cred = current_cred();
Dave Hansen4a3fd212008-02-15 14:37:48 -0800611 struct file *result;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700612 int ret;
613
Al Viro564f6992008-12-14 04:02:26 -0500614 if (attr) {
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300615 if (!mq_attr_ok(ipc_ns, attr)) {
616 ret = -EINVAL;
Alexander Viro7c7dce92006-01-14 15:29:55 -0500617 goto out;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300618 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619 /* store for use during create */
Al Viro564f6992008-12-14 04:02:26 -0500620 dentry->d_fsdata = attr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700621 }
622
Al Viroce3b0f82009-03-29 19:08:22 -0400623 mode &= ~current_umask();
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700624 ret = mnt_want_write(ipc_ns->mq_mnt);
Dave Hansen4a3fd212008-02-15 14:37:48 -0800625 if (ret)
626 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627 ret = vfs_create(dir->d_inode, dentry, mode, NULL);
628 dentry->d_fsdata = NULL;
629 if (ret)
Dave Hansen4a3fd212008-02-15 14:37:48 -0800630 goto out_drop_write;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700631
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700632 result = dentry_open(dentry, ipc_ns->mq_mnt, oflag, cred);
Dave Hansen4a3fd212008-02-15 14:37:48 -0800633 /*
634 * dentry_open() took a persistent mnt_want_write(),
635 * so we can now drop this one.
636 */
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700637 mnt_drop_write(ipc_ns->mq_mnt);
Dave Hansen4a3fd212008-02-15 14:37:48 -0800638 return result;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639
Dave Hansen4a3fd212008-02-15 14:37:48 -0800640out_drop_write:
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700641 mnt_drop_write(ipc_ns->mq_mnt);
Alexander Viro7c7dce92006-01-14 15:29:55 -0500642out:
643 dput(dentry);
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700644 mntput(ipc_ns->mq_mnt);
Alexander Viro7c7dce92006-01-14 15:29:55 -0500645 return ERR_PTR(ret);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700646}
647
648/* Opens existing queue */
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700649static struct file *do_open(struct ipc_namespace *ipc_ns,
650 struct dentry *dentry, int oflag)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651{
André Goddard Rosa04db0dd2010-02-23 04:04:25 -0300652 int ret;
David Howells745ca242008-11-14 10:39:22 +1100653 const struct cred *cred = current_cred();
654
655 static const int oflag2acc[O_ACCMODE] = { MAY_READ, MAY_WRITE,
656 MAY_READ | MAY_WRITE };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657
Alexander Viro7c7dce92006-01-14 15:29:55 -0500658 if ((oflag & O_ACCMODE) == (O_RDWR | O_WRONLY)) {
André Goddard Rosa04db0dd2010-02-23 04:04:25 -0300659 ret = -EINVAL;
660 goto err;
Alexander Viro7c7dce92006-01-14 15:29:55 -0500661 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662
Al Virof419a2e2008-07-22 00:07:17 -0400663 if (inode_permission(dentry->d_inode, oflag2acc[oflag & O_ACCMODE])) {
André Goddard Rosa04db0dd2010-02-23 04:04:25 -0300664 ret = -EACCES;
665 goto err;
Alexander Viro7c7dce92006-01-14 15:29:55 -0500666 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700668 return dentry_open(dentry, ipc_ns->mq_mnt, oflag, cred);
André Goddard Rosa04db0dd2010-02-23 04:04:25 -0300669
670err:
671 dput(dentry);
672 mntput(ipc_ns->mq_mnt);
673 return ERR_PTR(ret);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674}
675
Al Virodf0a4282011-07-26 05:26:10 -0400676SYSCALL_DEFINE4(mq_open, const char __user *, u_name, int, oflag, umode_t, mode,
Heiko Carstensd5460c92009-01-14 14:14:27 +0100677 struct mq_attr __user *, u_attr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678{
679 struct dentry *dentry;
680 struct file *filp;
681 char *name;
Al Viro564f6992008-12-14 04:02:26 -0500682 struct mq_attr attr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683 int fd, error;
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700684 struct ipc_namespace *ipc_ns = current->nsproxy->ipc_ns;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700685
Al Viro564f6992008-12-14 04:02:26 -0500686 if (u_attr && copy_from_user(&attr, u_attr, sizeof(struct mq_attr)))
687 return -EFAULT;
688
689 audit_mq_open(oflag, mode, u_attr ? &attr : NULL);
George C. Wilson20ca73b2006-05-24 16:09:55 -0500690
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691 if (IS_ERR(name = getname(u_name)))
692 return PTR_ERR(name);
693
Ulrich Drepper269f2132008-05-03 15:28:45 -0400694 fd = get_unused_fd_flags(O_CLOEXEC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695 if (fd < 0)
696 goto out_putname;
697
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700698 mutex_lock(&ipc_ns->mq_mnt->mnt_root->d_inode->i_mutex);
699 dentry = lookup_one_len(name, ipc_ns->mq_mnt->mnt_root, strlen(name));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700 if (IS_ERR(dentry)) {
701 error = PTR_ERR(dentry);
André Goddard Rosa4294a8e2010-02-23 04:04:28 -0300702 goto out_putfd;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703 }
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700704 mntget(ipc_ns->mq_mnt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705
706 if (oflag & O_CREAT) {
707 if (dentry->d_inode) { /* entry already exists */
Al Viro5a190ae2007-06-07 12:19:32 -0400708 audit_inode(name, dentry);
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300709 if (oflag & O_EXCL) {
710 error = -EEXIST;
Alexander Viro7c7dce92006-01-14 15:29:55 -0500711 goto out;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300712 }
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700713 filp = do_open(ipc_ns, dentry, oflag);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714 } else {
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700715 filp = do_create(ipc_ns, ipc_ns->mq_mnt->mnt_root,
716 dentry, oflag, mode,
Al Viro564f6992008-12-14 04:02:26 -0500717 u_attr ? &attr : NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700718 }
Alexander Viro7c7dce92006-01-14 15:29:55 -0500719 } else {
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300720 if (!dentry->d_inode) {
721 error = -ENOENT;
Alexander Viro7c7dce92006-01-14 15:29:55 -0500722 goto out;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300723 }
Al Viro5a190ae2007-06-07 12:19:32 -0400724 audit_inode(name, dentry);
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700725 filp = do_open(ipc_ns, dentry, oflag);
Alexander Viro7c7dce92006-01-14 15:29:55 -0500726 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727
728 if (IS_ERR(filp)) {
729 error = PTR_ERR(filp);
730 goto out_putfd;
731 }
732
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733 fd_install(fd, filp);
734 goto out_upsem;
735
Alexander Viro7c7dce92006-01-14 15:29:55 -0500736out:
737 dput(dentry);
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700738 mntput(ipc_ns->mq_mnt);
Alexander Viro7c7dce92006-01-14 15:29:55 -0500739out_putfd:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740 put_unused_fd(fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741 fd = error;
742out_upsem:
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700743 mutex_unlock(&ipc_ns->mq_mnt->mnt_root->d_inode->i_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744out_putname:
745 putname(name);
746 return fd;
747}
748
Heiko Carstensd5460c92009-01-14 14:14:27 +0100749SYSCALL_DEFINE1(mq_unlink, const char __user *, u_name)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700750{
751 int err;
752 char *name;
753 struct dentry *dentry;
754 struct inode *inode = NULL;
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700755 struct ipc_namespace *ipc_ns = current->nsproxy->ipc_ns;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700756
757 name = getname(u_name);
758 if (IS_ERR(name))
759 return PTR_ERR(name);
760
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700761 mutex_lock_nested(&ipc_ns->mq_mnt->mnt_root->d_inode->i_mutex,
Peter Zijlstra7a434812007-03-06 01:42:09 -0800762 I_MUTEX_PARENT);
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700763 dentry = lookup_one_len(name, ipc_ns->mq_mnt->mnt_root, strlen(name));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700764 if (IS_ERR(dentry)) {
765 err = PTR_ERR(dentry);
766 goto out_unlock;
767 }
768
769 if (!dentry->d_inode) {
770 err = -ENOENT;
771 goto out_err;
772 }
773
774 inode = dentry->d_inode;
775 if (inode)
Al Viro7de9c6e2010-10-23 11:11:40 -0400776 ihold(inode);
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700777 err = mnt_want_write(ipc_ns->mq_mnt);
Dave Hansen06227532008-02-15 14:37:34 -0800778 if (err)
779 goto out_err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780 err = vfs_unlink(dentry->d_parent->d_inode, dentry);
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700781 mnt_drop_write(ipc_ns->mq_mnt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700782out_err:
783 dput(dentry);
784
785out_unlock:
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700786 mutex_unlock(&ipc_ns->mq_mnt->mnt_root->d_inode->i_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787 putname(name);
788 if (inode)
789 iput(inode);
790
791 return err;
792}
793
794/* Pipelined send and receive functions.
795 *
796 * If a receiver finds no waiting message, then it registers itself in the
797 * list of waiting receivers. A sender checks that list before adding the new
798 * message into the message array. If there is a waiting receiver, then it
799 * bypasses the message array and directly hands the message over to the
800 * receiver.
801 * The receiver accepts the message and returns without grabbing the queue
802 * spinlock. Therefore an intermediate STATE_PENDING state and memory barriers
803 * are necessary. The same algorithm is used for sysv semaphores, see
Serge E. Hallyn7b7a3172006-03-28 01:56:23 -0800804 * ipc/sem.c for more details.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805 *
806 * The same algorithm is used for senders.
807 */
808
809/* pipelined_send() - send a message directly to the task waiting in
810 * sys_mq_timedreceive() (without inserting message into a queue).
811 */
812static inline void pipelined_send(struct mqueue_inode_info *info,
813 struct msg_msg *message,
814 struct ext_wait_queue *receiver)
815{
816 receiver->msg = message;
817 list_del(&receiver->list);
818 receiver->state = STATE_PENDING;
819 wake_up_process(receiver->task);
akpm@osdl.orgd59dd462005-05-01 08:58:47 -0700820 smp_wmb();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700821 receiver->state = STATE_READY;
822}
823
824/* pipelined_receive() - if there is task waiting in sys_mq_timedsend()
825 * gets its message and put to the queue (we have one free place for sure). */
826static inline void pipelined_receive(struct mqueue_inode_info *info)
827{
828 struct ext_wait_queue *sender = wq_get_first_waiter(info, SEND);
829
830 if (!sender) {
831 /* for poll */
832 wake_up_interruptible(&info->wait_q);
833 return;
834 }
835 msg_insert(sender->msg, info);
836 list_del(&sender->list);
837 sender->state = STATE_PENDING;
838 wake_up_process(sender->task);
akpm@osdl.orgd59dd462005-05-01 08:58:47 -0700839 smp_wmb();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700840 sender->state = STATE_READY;
841}
842
Heiko Carstensc4ea37c2009-01-14 14:14:28 +0100843SYSCALL_DEFINE5(mq_timedsend, mqd_t, mqdes, const char __user *, u_msg_ptr,
844 size_t, msg_len, unsigned int, msg_prio,
845 const struct timespec __user *, u_abs_timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700846{
847 struct file *filp;
848 struct inode *inode;
849 struct ext_wait_queue wait;
850 struct ext_wait_queue *receiver;
851 struct msg_msg *msg_ptr;
852 struct mqueue_inode_info *info;
Carsten Emde9ca7d8e2010-04-02 22:40:20 +0200853 ktime_t expires, *timeout = NULL;
854 struct timespec ts;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855 int ret;
856
Al Viroc32c8af2008-12-14 03:46:48 -0500857 if (u_abs_timeout) {
Carsten Emde9ca7d8e2010-04-02 22:40:20 +0200858 int res = prepare_timeout(u_abs_timeout, &expires, &ts);
859 if (res)
860 return res;
861 timeout = &expires;
Al Viroc32c8af2008-12-14 03:46:48 -0500862 }
George C. Wilson20ca73b2006-05-24 16:09:55 -0500863
Linus Torvalds1da177e2005-04-16 15:20:36 -0700864 if (unlikely(msg_prio >= (unsigned long) MQ_PRIO_MAX))
865 return -EINVAL;
866
Carsten Emde9ca7d8e2010-04-02 22:40:20 +0200867 audit_mq_sendrecv(mqdes, msg_len, msg_prio, timeout ? &ts : NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700868
Linus Torvalds1da177e2005-04-16 15:20:36 -0700869 filp = fget(mqdes);
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300870 if (unlikely(!filp)) {
871 ret = -EBADF;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700872 goto out;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300873 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700874
Josef Sipek6d630792006-12-08 02:37:11 -0800875 inode = filp->f_path.dentry->d_inode;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300876 if (unlikely(filp->f_op != &mqueue_file_operations)) {
877 ret = -EBADF;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700878 goto out_fput;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300879 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700880 info = MQUEUE_I(inode);
Al Viro5a190ae2007-06-07 12:19:32 -0400881 audit_inode(NULL, filp->f_path.dentry);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700882
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300883 if (unlikely(!(filp->f_mode & FMODE_WRITE))) {
884 ret = -EBADF;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885 goto out_fput;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300886 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700887
888 if (unlikely(msg_len > info->attr.mq_msgsize)) {
889 ret = -EMSGSIZE;
890 goto out_fput;
891 }
892
893 /* First try to allocate memory, before doing anything with
894 * existing queues. */
895 msg_ptr = load_msg(u_msg_ptr, msg_len);
896 if (IS_ERR(msg_ptr)) {
897 ret = PTR_ERR(msg_ptr);
898 goto out_fput;
899 }
900 msg_ptr->m_ts = msg_len;
901 msg_ptr->m_type = msg_prio;
902
903 spin_lock(&info->lock);
904
905 if (info->attr.mq_curmsgs == info->attr.mq_maxmsg) {
906 if (filp->f_flags & O_NONBLOCK) {
907 spin_unlock(&info->lock);
908 ret = -EAGAIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909 } else {
910 wait.task = current;
911 wait.msg = (void *) msg_ptr;
912 wait.state = STATE_NONE;
913 ret = wq_sleep(info, SEND, timeout, &wait);
914 }
915 if (ret < 0)
916 free_msg(msg_ptr);
917 } else {
918 receiver = wq_get_first_waiter(info, RECV);
919 if (receiver) {
920 pipelined_send(info, msg_ptr, receiver);
921 } else {
922 /* adds message to the queue */
923 msg_insert(msg_ptr, info);
924 __do_notify(info);
925 }
926 inode->i_atime = inode->i_mtime = inode->i_ctime =
927 CURRENT_TIME;
928 spin_unlock(&info->lock);
929 ret = 0;
930 }
931out_fput:
932 fput(filp);
933out:
934 return ret;
935}
936
Heiko Carstensc4ea37c2009-01-14 14:14:28 +0100937SYSCALL_DEFINE5(mq_timedreceive, mqd_t, mqdes, char __user *, u_msg_ptr,
938 size_t, msg_len, unsigned int __user *, u_msg_prio,
939 const struct timespec __user *, u_abs_timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700940{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700941 ssize_t ret;
942 struct msg_msg *msg_ptr;
943 struct file *filp;
944 struct inode *inode;
945 struct mqueue_inode_info *info;
946 struct ext_wait_queue wait;
Carsten Emde9ca7d8e2010-04-02 22:40:20 +0200947 ktime_t expires, *timeout = NULL;
948 struct timespec ts;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949
Al Viroc32c8af2008-12-14 03:46:48 -0500950 if (u_abs_timeout) {
Carsten Emde9ca7d8e2010-04-02 22:40:20 +0200951 int res = prepare_timeout(u_abs_timeout, &expires, &ts);
952 if (res)
953 return res;
954 timeout = &expires;
Al Viroc32c8af2008-12-14 03:46:48 -0500955 }
George C. Wilson20ca73b2006-05-24 16:09:55 -0500956
Carsten Emde9ca7d8e2010-04-02 22:40:20 +0200957 audit_mq_sendrecv(mqdes, msg_len, 0, timeout ? &ts : NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959 filp = fget(mqdes);
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300960 if (unlikely(!filp)) {
961 ret = -EBADF;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962 goto out;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300963 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700964
Josef Sipek6d630792006-12-08 02:37:11 -0800965 inode = filp->f_path.dentry->d_inode;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300966 if (unlikely(filp->f_op != &mqueue_file_operations)) {
967 ret = -EBADF;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700968 goto out_fput;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300969 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700970 info = MQUEUE_I(inode);
Al Viro5a190ae2007-06-07 12:19:32 -0400971 audit_inode(NULL, filp->f_path.dentry);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700972
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300973 if (unlikely(!(filp->f_mode & FMODE_READ))) {
974 ret = -EBADF;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700975 goto out_fput;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300976 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977
978 /* checks if buffer is big enough */
979 if (unlikely(msg_len < info->attr.mq_msgsize)) {
980 ret = -EMSGSIZE;
981 goto out_fput;
982 }
983
984 spin_lock(&info->lock);
985 if (info->attr.mq_curmsgs == 0) {
986 if (filp->f_flags & O_NONBLOCK) {
987 spin_unlock(&info->lock);
988 ret = -EAGAIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700989 } else {
990 wait.task = current;
991 wait.state = STATE_NONE;
992 ret = wq_sleep(info, RECV, timeout, &wait);
993 msg_ptr = wait.msg;
994 }
995 } else {
996 msg_ptr = msg_get(info);
997
998 inode->i_atime = inode->i_mtime = inode->i_ctime =
999 CURRENT_TIME;
1000
1001 /* There is now free space in queue. */
1002 pipelined_receive(info);
1003 spin_unlock(&info->lock);
1004 ret = 0;
1005 }
1006 if (ret == 0) {
1007 ret = msg_ptr->m_ts;
1008
1009 if ((u_msg_prio && put_user(msg_ptr->m_type, u_msg_prio)) ||
1010 store_msg(u_msg_ptr, msg_ptr, msg_ptr->m_ts)) {
1011 ret = -EFAULT;
1012 }
1013 free_msg(msg_ptr);
1014 }
1015out_fput:
1016 fput(filp);
1017out:
1018 return ret;
1019}
1020
1021/*
1022 * Notes: the case when user wants us to deregister (with NULL as pointer)
1023 * and he isn't currently owner of notification, will be silently discarded.
1024 * It isn't explicitly defined in the POSIX.
1025 */
Heiko Carstensc4ea37c2009-01-14 14:14:28 +01001026SYSCALL_DEFINE2(mq_notify, mqd_t, mqdes,
1027 const struct sigevent __user *, u_notification)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001028{
1029 int ret;
1030 struct file *filp;
1031 struct sock *sock;
1032 struct inode *inode;
1033 struct sigevent notification;
1034 struct mqueue_inode_info *info;
1035 struct sk_buff *nc;
1036
Al Viro20114f72008-12-10 07:16:12 -05001037 if (u_notification) {
1038 if (copy_from_user(&notification, u_notification,
1039 sizeof(struct sigevent)))
1040 return -EFAULT;
1041 }
1042
1043 audit_mq_notify(mqdes, u_notification ? &notification : NULL);
George C. Wilson20ca73b2006-05-24 16:09:55 -05001044
Linus Torvalds1da177e2005-04-16 15:20:36 -07001045 nc = NULL;
1046 sock = NULL;
1047 if (u_notification != NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001048 if (unlikely(notification.sigev_notify != SIGEV_NONE &&
1049 notification.sigev_notify != SIGEV_SIGNAL &&
1050 notification.sigev_notify != SIGEV_THREAD))
1051 return -EINVAL;
1052 if (notification.sigev_notify == SIGEV_SIGNAL &&
Jesper Juhl7ed20e12005-05-01 08:59:14 -07001053 !valid_signal(notification.sigev_signo)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001054 return -EINVAL;
1055 }
1056 if (notification.sigev_notify == SIGEV_THREAD) {
Patrick McHardyc3d8d1e2007-11-07 02:42:09 -08001057 long timeo;
1058
Linus Torvalds1da177e2005-04-16 15:20:36 -07001059 /* create the notify skb */
1060 nc = alloc_skb(NOTIFY_COOKIE_LEN, GFP_KERNEL);
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -03001061 if (!nc) {
1062 ret = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001063 goto out;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -03001064 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001065 if (copy_from_user(nc->data,
1066 notification.sigev_value.sival_ptr,
1067 NOTIFY_COOKIE_LEN)) {
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -03001068 ret = -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001069 goto out;
1070 }
1071
1072 /* TODO: add a header? */
1073 skb_put(nc, NOTIFY_COOKIE_LEN);
1074 /* and attach it to the socket */
1075retry:
1076 filp = fget(notification.sigev_signo);
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -03001077 if (!filp) {
1078 ret = -EBADF;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001079 goto out;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -03001080 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001081 sock = netlink_getsockbyfilp(filp);
1082 fput(filp);
1083 if (IS_ERR(sock)) {
1084 ret = PTR_ERR(sock);
1085 sock = NULL;
1086 goto out;
1087 }
1088
Patrick McHardyc3d8d1e2007-11-07 02:42:09 -08001089 timeo = MAX_SCHEDULE_TIMEOUT;
Denis V. Lunev9457afe2008-06-05 11:23:39 -07001090 ret = netlink_attachskb(sock, nc, &timeo, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001091 if (ret == 1)
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -03001092 goto retry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001093 if (ret) {
1094 sock = NULL;
1095 nc = NULL;
1096 goto out;
1097 }
1098 }
1099 }
1100
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101 filp = fget(mqdes);
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -03001102 if (!filp) {
1103 ret = -EBADF;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001104 goto out;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -03001105 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001106
Josef Sipek6d630792006-12-08 02:37:11 -08001107 inode = filp->f_path.dentry->d_inode;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -03001108 if (unlikely(filp->f_op != &mqueue_file_operations)) {
1109 ret = -EBADF;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001110 goto out_fput;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -03001111 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112 info = MQUEUE_I(inode);
1113
1114 ret = 0;
1115 spin_lock(&info->lock);
1116 if (u_notification == NULL) {
Cedric Le Goatera03fcb72006-10-02 02:17:26 -07001117 if (info->notify_owner == task_tgid(current)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001118 remove_notification(info);
1119 inode->i_atime = inode->i_ctime = CURRENT_TIME;
1120 }
Cedric Le Goatera03fcb72006-10-02 02:17:26 -07001121 } else if (info->notify_owner != NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001122 ret = -EBUSY;
1123 } else {
1124 switch (notification.sigev_notify) {
1125 case SIGEV_NONE:
1126 info->notify.sigev_notify = SIGEV_NONE;
1127 break;
1128 case SIGEV_THREAD:
1129 info->notify_sock = sock;
1130 info->notify_cookie = nc;
1131 sock = NULL;
1132 nc = NULL;
1133 info->notify.sigev_notify = SIGEV_THREAD;
1134 break;
1135 case SIGEV_SIGNAL:
1136 info->notify.sigev_signo = notification.sigev_signo;
1137 info->notify.sigev_value = notification.sigev_value;
1138 info->notify.sigev_notify = SIGEV_SIGNAL;
1139 break;
1140 }
Cedric Le Goatera03fcb72006-10-02 02:17:26 -07001141
1142 info->notify_owner = get_pid(task_tgid(current));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143 inode->i_atime = inode->i_ctime = CURRENT_TIME;
1144 }
1145 spin_unlock(&info->lock);
1146out_fput:
1147 fput(filp);
1148out:
1149 if (sock) {
1150 netlink_detachskb(sock, nc);
1151 } else if (nc) {
1152 dev_kfree_skb(nc);
1153 }
1154 return ret;
1155}
1156
Heiko Carstensc4ea37c2009-01-14 14:14:28 +01001157SYSCALL_DEFINE3(mq_getsetattr, mqd_t, mqdes,
1158 const struct mq_attr __user *, u_mqstat,
1159 struct mq_attr __user *, u_omqstat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001160{
1161 int ret;
1162 struct mq_attr mqstat, omqstat;
1163 struct file *filp;
1164 struct inode *inode;
1165 struct mqueue_inode_info *info;
1166
1167 if (u_mqstat != NULL) {
1168 if (copy_from_user(&mqstat, u_mqstat, sizeof(struct mq_attr)))
1169 return -EFAULT;
1170 if (mqstat.mq_flags & (~O_NONBLOCK))
1171 return -EINVAL;
1172 }
1173
Linus Torvalds1da177e2005-04-16 15:20:36 -07001174 filp = fget(mqdes);
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -03001175 if (!filp) {
1176 ret = -EBADF;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001177 goto out;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -03001178 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001179
Josef Sipek6d630792006-12-08 02:37:11 -08001180 inode = filp->f_path.dentry->d_inode;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -03001181 if (unlikely(filp->f_op != &mqueue_file_operations)) {
1182 ret = -EBADF;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001183 goto out_fput;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -03001184 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001185 info = MQUEUE_I(inode);
1186
1187 spin_lock(&info->lock);
1188
1189 omqstat = info->attr;
1190 omqstat.mq_flags = filp->f_flags & O_NONBLOCK;
1191 if (u_mqstat) {
Al Viro73929062008-12-10 06:58:59 -05001192 audit_mq_getsetattr(mqdes, &mqstat);
Jonathan Corbetdb1dd4d2009-02-06 15:25:24 -07001193 spin_lock(&filp->f_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001194 if (mqstat.mq_flags & O_NONBLOCK)
1195 filp->f_flags |= O_NONBLOCK;
1196 else
1197 filp->f_flags &= ~O_NONBLOCK;
Jonathan Corbetdb1dd4d2009-02-06 15:25:24 -07001198 spin_unlock(&filp->f_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001199
1200 inode->i_atime = inode->i_ctime = CURRENT_TIME;
1201 }
1202
1203 spin_unlock(&info->lock);
1204
1205 ret = 0;
1206 if (u_omqstat != NULL && copy_to_user(u_omqstat, &omqstat,
1207 sizeof(struct mq_attr)))
1208 ret = -EFAULT;
1209
1210out_fput:
1211 fput(filp);
1212out:
1213 return ret;
1214}
1215
Arjan van de Ven92e1d5b2007-02-12 00:55:39 -08001216static const struct inode_operations mqueue_dir_inode_operations = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001217 .lookup = simple_lookup,
1218 .create = mqueue_create,
1219 .unlink = mqueue_unlink,
1220};
1221
Arjan van de Ven9a321442007-02-12 00:55:35 -08001222static const struct file_operations mqueue_file_operations = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001223 .flush = mqueue_flush_file,
1224 .poll = mqueue_poll_file,
1225 .read = mqueue_read_file,
Arnd Bergmann6038f372010-08-15 18:52:59 +02001226 .llseek = default_llseek,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001227};
1228
Alexey Dobriyanb87221d2009-09-21 17:01:09 -07001229static const struct super_operations mqueue_super_ops = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001230 .alloc_inode = mqueue_alloc_inode,
1231 .destroy_inode = mqueue_destroy_inode,
Al Viro6d8af642010-06-05 16:29:45 -04001232 .evict_inode = mqueue_evict_inode,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001233 .statfs = simple_statfs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001234};
1235
1236static struct file_system_type mqueue_fs_type = {
1237 .name = "mqueue",
Al Viroceefda62010-07-26 13:16:50 +04001238 .mount = mqueue_mount,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001239 .kill_sb = kill_litter_super,
1240};
1241
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -07001242int mq_init_ns(struct ipc_namespace *ns)
1243{
1244 ns->mq_queues_count = 0;
1245 ns->mq_queues_max = DFLT_QUEUESMAX;
1246 ns->mq_msg_max = DFLT_MSGMAX;
1247 ns->mq_msgsize_max = DFLT_MSGSIZEMAX;
1248
1249 ns->mq_mnt = kern_mount_data(&mqueue_fs_type, ns);
1250 if (IS_ERR(ns->mq_mnt)) {
1251 int err = PTR_ERR(ns->mq_mnt);
1252 ns->mq_mnt = NULL;
1253 return err;
1254 }
1255 return 0;
1256}
1257
1258void mq_clear_sbinfo(struct ipc_namespace *ns)
1259{
1260 ns->mq_mnt->mnt_sb->s_fs_info = NULL;
1261}
1262
1263void mq_put_mnt(struct ipc_namespace *ns)
1264{
Al Viro6f686572011-12-09 00:38:50 -05001265 kern_unmount(ns->mq_mnt);
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -07001266}
1267
Linus Torvalds1da177e2005-04-16 15:20:36 -07001268static int __init init_mqueue_fs(void)
1269{
1270 int error;
1271
1272 mqueue_inode_cachep = kmem_cache_create("mqueue_inode_cache",
1273 sizeof(struct mqueue_inode_info), 0,
Paul Mundt20c2df82007-07-20 10:11:58 +09001274 SLAB_HWCACHE_ALIGN, init_once);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275 if (mqueue_inode_cachep == NULL)
1276 return -ENOMEM;
1277
André Goddard Rosa2329e392010-02-23 04:04:27 -03001278 /* ignore failures - they are not fatal */
Serge E. Hallynbdc8e5f2009-04-06 19:01:11 -07001279 mq_sysctl_table = mq_register_sysctl_table();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001280
1281 error = register_filesystem(&mqueue_fs_type);
1282 if (error)
1283 goto out_sysctl;
1284
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -07001285 spin_lock_init(&mq_lock);
1286
Al Viro6f686572011-12-09 00:38:50 -05001287 error = mq_init_ns(&init_ipc_ns);
1288 if (error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289 goto out_filesystem;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001290
Linus Torvalds1da177e2005-04-16 15:20:36 -07001291 return 0;
1292
1293out_filesystem:
1294 unregister_filesystem(&mqueue_fs_type);
1295out_sysctl:
1296 if (mq_sysctl_table)
1297 unregister_sysctl_table(mq_sysctl_table);
Alexey Dobriyan1a1d92c2006-09-27 01:49:40 -07001298 kmem_cache_destroy(mqueue_inode_cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001299 return error;
1300}
1301
1302__initcall(init_mqueue_fs);