blob: ed049ea568f45ee6026941b5d8335853c33efcd8 [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>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090035#include <linux/slab.h>
Ingo Molnar5f921ae2006-03-26 01:37:17 -080036
Linus Torvalds1da177e2005-04-16 15:20:36 -070037#include <net/sock.h>
38#include "util.h"
39
40#define MQUEUE_MAGIC 0x19800202
41#define DIRENT_SIZE 20
42#define FILENT_SIZE 80
43
44#define SEND 0
45#define RECV 1
46
47#define STATE_NONE 0
48#define STATE_PENDING 1
49#define STATE_READY 2
50
Linus Torvalds1da177e2005-04-16 15:20:36 -070051struct ext_wait_queue { /* queue of sleeping tasks */
52 struct task_struct *task;
53 struct list_head list;
54 struct msg_msg *msg; /* ptr of loaded message */
55 int state; /* one of STATE_* values */
56};
57
58struct mqueue_inode_info {
59 spinlock_t lock;
60 struct inode vfs_inode;
61 wait_queue_head_t wait_q;
62
63 struct msg_msg **messages;
64 struct mq_attr attr;
65
66 struct sigevent notify;
Cedric Le Goatera03fcb72006-10-02 02:17:26 -070067 struct pid* notify_owner;
Adrian Bunk338cec32005-09-10 00:26:54 -070068 struct user_struct *user; /* user who created, for accounting */
Linus Torvalds1da177e2005-04-16 15:20:36 -070069 struct sock *notify_sock;
70 struct sk_buff *notify_cookie;
71
72 /* for tasks waiting for free space and messages, respectively */
73 struct ext_wait_queue e_wait_q[2];
74
75 unsigned long qsize; /* size of queue in memory (sum of all msgs) */
76};
77
Arjan van de Ven92e1d5b2007-02-12 00:55:39 -080078static const struct inode_operations mqueue_dir_inode_operations;
Arjan van de Ven9a321442007-02-12 00:55:35 -080079static const struct file_operations mqueue_file_operations;
Alexey Dobriyanb87221d2009-09-21 17:01:09 -070080static const struct super_operations mqueue_super_ops;
Linus Torvalds1da177e2005-04-16 15:20:36 -070081static void remove_notification(struct mqueue_inode_info *info);
82
Christoph Lametere18b8902006-12-06 20:33:20 -080083static struct kmem_cache *mqueue_inode_cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -070084
85static struct ctl_table_header * mq_sysctl_table;
86
87static inline struct mqueue_inode_info *MQUEUE_I(struct inode *inode)
88{
89 return container_of(inode, struct mqueue_inode_info, vfs_inode);
90}
91
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -070092/*
93 * This routine should be called with the mq_lock held.
94 */
95static inline struct ipc_namespace *__get_ns_from_inode(struct inode *inode)
Serge E. Hallyn614b84c2009-04-06 19:01:08 -070096{
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -070097 return get_ipc_ns(inode->i_sb->s_fs_info);
Serge E. Hallyn614b84c2009-04-06 19:01:08 -070098}
99
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700100static struct ipc_namespace *get_ns_from_inode(struct inode *inode)
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700101{
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700102 struct ipc_namespace *ns;
103
104 spin_lock(&mq_lock);
105 ns = __get_ns_from_inode(inode);
106 spin_unlock(&mq_lock);
107 return ns;
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700108}
109
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700110static struct inode *mqueue_get_inode(struct super_block *sb,
111 struct ipc_namespace *ipc_ns, int mode,
112 struct mq_attr *attr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113{
David Howells86a264a2008-11-14 10:39:18 +1100114 struct user_struct *u = current_user();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700115 struct inode *inode;
Jiri Slabyd40dcdb2011-07-26 16:08:47 -0700116 int ret = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700117
118 inode = new_inode(sb);
Jiri Slaby04715202011-07-26 16:08:46 -0700119 if (!inode)
120 goto err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700121
Jiri Slaby04715202011-07-26 16:08:46 -0700122 inode->i_ino = get_next_ino();
123 inode->i_mode = mode;
124 inode->i_uid = current_fsuid();
125 inode->i_gid = current_fsgid();
126 inode->i_mtime = inode->i_ctime = inode->i_atime = CURRENT_TIME;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700127
Jiri Slaby04715202011-07-26 16:08:46 -0700128 if (S_ISREG(mode)) {
129 struct mqueue_inode_info *info;
130 struct task_struct *p = current;
131 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 ||
161 u->mq_bytes + mq_bytes > task_rlimit(p, RLIMIT_MSGQUEUE)) {
162 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;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300191 int error;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192
193 sb->s_blocksize = PAGE_CACHE_SIZE;
194 sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
195 sb->s_magic = MQUEUE_MAGIC;
196 sb->s_op = &mqueue_super_ops;
197
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700198 inode = mqueue_get_inode(sb, ns, S_IFDIR | S_ISVTX | S_IRWXUGO,
199 NULL);
Jiri Slabyd40dcdb2011-07-26 16:08:47 -0700200 if (IS_ERR(inode)) {
201 error = PTR_ERR(inode);
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700202 goto out;
203 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700204
205 sb->s_root = d_alloc_root(inode);
206 if (!sb->s_root) {
207 iput(inode);
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700208 error = -ENOMEM;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300209 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700210 }
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300211 error = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700213out:
214 return error;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700215}
216
Al Viroceefda62010-07-26 13:16:50 +0400217static struct dentry *mqueue_mount(struct file_system_type *fs_type,
David Howells454e2392006-06-23 02:02:57 -0700218 int flags, const char *dev_name,
Al Viroceefda62010-07-26 13:16:50 +0400219 void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700220{
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700221 if (!(flags & MS_KERNMOUNT))
222 data = current->nsproxy->ipc_ns;
Al Viroceefda62010-07-26 13:16:50 +0400223 return mount_ns(fs_type, flags, data, mqueue_fill_super);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700224}
225
Alexey Dobriyan51cc5062008-07-25 19:45:34 -0700226static void init_once(void *foo)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700227{
228 struct mqueue_inode_info *p = (struct mqueue_inode_info *) foo;
229
Christoph Lametera35afb82007-05-16 22:10:57 -0700230 inode_init_once(&p->vfs_inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231}
232
233static struct inode *mqueue_alloc_inode(struct super_block *sb)
234{
235 struct mqueue_inode_info *ei;
236
Christoph Lametere94b1762006-12-06 20:33:17 -0800237 ei = kmem_cache_alloc(mqueue_inode_cachep, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700238 if (!ei)
239 return NULL;
240 return &ei->vfs_inode;
241}
242
Nick Pigginfa0d7e32011-01-07 17:49:49 +1100243static void mqueue_i_callback(struct rcu_head *head)
244{
245 struct inode *inode = container_of(head, struct inode, i_rcu);
246 INIT_LIST_HEAD(&inode->i_dentry);
247 kmem_cache_free(mqueue_inode_cachep, MQUEUE_I(inode));
248}
249
Linus Torvalds1da177e2005-04-16 15:20:36 -0700250static void mqueue_destroy_inode(struct inode *inode)
251{
Nick Pigginfa0d7e32011-01-07 17:49:49 +1100252 call_rcu(&inode->i_rcu, mqueue_i_callback);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253}
254
Al Viro6d8af642010-06-05 16:29:45 -0400255static void mqueue_evict_inode(struct inode *inode)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700256{
257 struct mqueue_inode_info *info;
258 struct user_struct *user;
259 unsigned long mq_bytes;
260 int i;
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700261 struct ipc_namespace *ipc_ns;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700262
Al Viro6d8af642010-06-05 16:29:45 -0400263 end_writeback(inode);
264
265 if (S_ISDIR(inode->i_mode))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700266 return;
Al Viro6d8af642010-06-05 16:29:45 -0400267
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700268 ipc_ns = get_ns_from_inode(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269 info = MQUEUE_I(inode);
270 spin_lock(&info->lock);
271 for (i = 0; i < info->attr.mq_curmsgs; i++)
272 free_msg(info->messages[i]);
273 kfree(info->messages);
274 spin_unlock(&info->lock);
275
André Goddard Rosa8834cf72010-02-23 04:04:24 -0300276 /* Total amount of bytes accounted for the mqueue */
277 mq_bytes = info->attr.mq_maxmsg * (sizeof(struct msg_msg *)
278 + info->attr.mq_msgsize);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279 user = info->user;
280 if (user) {
281 spin_lock(&mq_lock);
282 user->mq_bytes -= mq_bytes;
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700283 /*
284 * get_ns_from_inode() ensures that the
285 * (ipc_ns = sb->s_fs_info) is either a valid ipc_ns
286 * to which we now hold a reference, or it is NULL.
287 * We can't put it here under mq_lock, though.
288 */
289 if (ipc_ns)
290 ipc_ns->mq_queues_count--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291 spin_unlock(&mq_lock);
292 free_uid(user);
293 }
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700294 if (ipc_ns)
295 put_ipc_ns(ipc_ns);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296}
297
298static int mqueue_create(struct inode *dir, struct dentry *dentry,
299 int mode, struct nameidata *nd)
300{
301 struct inode *inode;
302 struct mq_attr *attr = dentry->d_fsdata;
303 int error;
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700304 struct ipc_namespace *ipc_ns;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700305
306 spin_lock(&mq_lock);
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700307 ipc_ns = __get_ns_from_inode(dir);
308 if (!ipc_ns) {
309 error = -EACCES;
310 goto out_unlock;
311 }
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700312 if (ipc_ns->mq_queues_count >= ipc_ns->mq_queues_max &&
313 !capable(CAP_SYS_RESOURCE)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700314 error = -ENOSPC;
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700315 goto out_unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700316 }
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700317 ipc_ns->mq_queues_count++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318 spin_unlock(&mq_lock);
319
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700320 inode = mqueue_get_inode(dir->i_sb, ipc_ns, mode, attr);
Jiri Slabyd40dcdb2011-07-26 16:08:47 -0700321 if (IS_ERR(inode)) {
322 error = PTR_ERR(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700323 spin_lock(&mq_lock);
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700324 ipc_ns->mq_queues_count--;
325 goto out_unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700326 }
327
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700328 put_ipc_ns(ipc_ns);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700329 dir->i_size += DIRENT_SIZE;
330 dir->i_ctime = dir->i_mtime = dir->i_atime = CURRENT_TIME;
331
332 d_instantiate(dentry, inode);
333 dget(dentry);
334 return 0;
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700335out_unlock:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700336 spin_unlock(&mq_lock);
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700337 if (ipc_ns)
338 put_ipc_ns(ipc_ns);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700339 return error;
340}
341
342static int mqueue_unlink(struct inode *dir, struct dentry *dentry)
343{
344 struct inode *inode = dentry->d_inode;
345
346 dir->i_ctime = dir->i_mtime = dir->i_atime = CURRENT_TIME;
347 dir->i_size -= DIRENT_SIZE;
Dave Hansen9a53c3a2006-09-30 23:29:03 -0700348 drop_nlink(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700349 dput(dentry);
350 return 0;
351}
352
353/*
354* This is routine for system read from queue file.
355* To avoid mess with doing here some sort of mq_receive we allow
356* to read only queue size & notification info (the only values
357* that are interesting from user point of view and aren't accessible
358* through std routines)
359*/
360static ssize_t mqueue_read_file(struct file *filp, char __user *u_data,
Akinobu Mitaf1a43f92008-07-25 01:48:07 -0700361 size_t count, loff_t *off)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700362{
Josef Sipek6d630792006-12-08 02:37:11 -0800363 struct mqueue_inode_info *info = MQUEUE_I(filp->f_path.dentry->d_inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364 char buffer[FILENT_SIZE];
Akinobu Mitaf1a43f92008-07-25 01:48:07 -0700365 ssize_t ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700366
367 spin_lock(&info->lock);
368 snprintf(buffer, sizeof(buffer),
369 "QSIZE:%-10lu NOTIFY:%-5d SIGNO:%-5d NOTIFY_PID:%-6d\n",
370 info->qsize,
371 info->notify_owner ? info->notify.sigev_notify : 0,
372 (info->notify_owner &&
373 info->notify.sigev_notify == SIGEV_SIGNAL) ?
374 info->notify.sigev_signo : 0,
Pavel Emelyanov6c5f3e72008-02-08 04:19:20 -0800375 pid_vnr(info->notify_owner));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376 spin_unlock(&info->lock);
377 buffer[sizeof(buffer)-1] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378
Akinobu Mitaf1a43f92008-07-25 01:48:07 -0700379 ret = simple_read_from_buffer(u_data, count, off, buffer,
380 strlen(buffer));
381 if (ret <= 0)
382 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700383
Josef Sipek6d630792006-12-08 02:37:11 -0800384 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 -0700385 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386}
387
Miklos Szeredi75e1fcc2006-06-23 02:05:12 -0700388static int mqueue_flush_file(struct file *filp, fl_owner_t id)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389{
Josef Sipek6d630792006-12-08 02:37:11 -0800390 struct mqueue_inode_info *info = MQUEUE_I(filp->f_path.dentry->d_inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391
392 spin_lock(&info->lock);
Cedric Le Goatera03fcb72006-10-02 02:17:26 -0700393 if (task_tgid(current) == info->notify_owner)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394 remove_notification(info);
395
396 spin_unlock(&info->lock);
397 return 0;
398}
399
400static unsigned int mqueue_poll_file(struct file *filp, struct poll_table_struct *poll_tab)
401{
Josef Sipek6d630792006-12-08 02:37:11 -0800402 struct mqueue_inode_info *info = MQUEUE_I(filp->f_path.dentry->d_inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403 int retval = 0;
404
405 poll_wait(filp, &info->wait_q, poll_tab);
406
407 spin_lock(&info->lock);
408 if (info->attr.mq_curmsgs)
409 retval = POLLIN | POLLRDNORM;
410
411 if (info->attr.mq_curmsgs < info->attr.mq_maxmsg)
412 retval |= POLLOUT | POLLWRNORM;
413 spin_unlock(&info->lock);
414
415 return retval;
416}
417
418/* Adds current to info->e_wait_q[sr] before element with smaller prio */
419static void wq_add(struct mqueue_inode_info *info, int sr,
420 struct ext_wait_queue *ewp)
421{
422 struct ext_wait_queue *walk;
423
424 ewp->task = current;
425
426 list_for_each_entry(walk, &info->e_wait_q[sr].list, list) {
427 if (walk->task->static_prio <= current->static_prio) {
428 list_add_tail(&ewp->list, &walk->list);
429 return;
430 }
431 }
432 list_add_tail(&ewp->list, &info->e_wait_q[sr].list);
433}
434
435/*
436 * Puts current task to sleep. Caller must hold queue lock. After return
437 * lock isn't held.
438 * sr: SEND or RECV
439 */
440static int wq_sleep(struct mqueue_inode_info *info, int sr,
Carsten Emde9ca7d8e2010-04-02 22:40:20 +0200441 ktime_t *timeout, struct ext_wait_queue *ewp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442{
443 int retval;
444 signed long time;
445
446 wq_add(info, sr, ewp);
447
448 for (;;) {
449 set_current_state(TASK_INTERRUPTIBLE);
450
451 spin_unlock(&info->lock);
Carsten Emde9ca7d8e2010-04-02 22:40:20 +0200452 time = schedule_hrtimeout_range_clock(timeout,
453 HRTIMER_MODE_ABS, 0, CLOCK_REALTIME);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454
455 while (ewp->state == STATE_PENDING)
456 cpu_relax();
457
458 if (ewp->state == STATE_READY) {
459 retval = 0;
460 goto out;
461 }
462 spin_lock(&info->lock);
463 if (ewp->state == STATE_READY) {
464 retval = 0;
465 goto out_unlock;
466 }
467 if (signal_pending(current)) {
468 retval = -ERESTARTSYS;
469 break;
470 }
471 if (time == 0) {
472 retval = -ETIMEDOUT;
473 break;
474 }
475 }
476 list_del(&ewp->list);
477out_unlock:
478 spin_unlock(&info->lock);
479out:
480 return retval;
481}
482
483/*
484 * Returns waiting task that should be serviced first or NULL if none exists
485 */
486static struct ext_wait_queue *wq_get_first_waiter(
487 struct mqueue_inode_info *info, int sr)
488{
489 struct list_head *ptr;
490
491 ptr = info->e_wait_q[sr].list.prev;
492 if (ptr == &info->e_wait_q[sr].list)
493 return NULL;
494 return list_entry(ptr, struct ext_wait_queue, list);
495}
496
497/* Auxiliary functions to manipulate messages' list */
498static void msg_insert(struct msg_msg *ptr, struct mqueue_inode_info *info)
499{
500 int k;
501
502 k = info->attr.mq_curmsgs - 1;
503 while (k >= 0 && info->messages[k]->m_type >= ptr->m_type) {
504 info->messages[k + 1] = info->messages[k];
505 k--;
506 }
507 info->attr.mq_curmsgs++;
508 info->qsize += ptr->m_ts;
509 info->messages[k + 1] = ptr;
510}
511
512static inline struct msg_msg *msg_get(struct mqueue_inode_info *info)
513{
514 info->qsize -= info->messages[--info->attr.mq_curmsgs]->m_ts;
515 return info->messages[info->attr.mq_curmsgs];
516}
517
518static inline void set_cookie(struct sk_buff *skb, char code)
519{
520 ((char*)skb->data)[NOTIFY_COOKIE_LEN-1] = code;
521}
522
523/*
524 * The next function is only to split too long sys_mq_timedsend
525 */
526static void __do_notify(struct mqueue_inode_info *info)
527{
528 /* notification
529 * invoked when there is registered process and there isn't process
530 * waiting synchronously for message AND state of queue changed from
531 * empty to not empty. Here we are sure that no one is waiting
532 * synchronously. */
533 if (info->notify_owner &&
534 info->attr.mq_curmsgs == 1) {
535 struct siginfo sig_i;
536 switch (info->notify.sigev_notify) {
537 case SIGEV_NONE:
538 break;
539 case SIGEV_SIGNAL:
540 /* sends signal */
541
542 sig_i.si_signo = info->notify.sigev_signo;
543 sig_i.si_errno = 0;
544 sig_i.si_code = SI_MESGQ;
545 sig_i.si_value = info->notify.sigev_value;
Sukadev Bhattiprolua6684992009-01-07 18:08:50 -0800546 sig_i.si_pid = task_tgid_nr_ns(current,
547 ns_of_pid(info->notify_owner));
David Howells414c0702008-11-14 10:39:06 +1100548 sig_i.si_uid = current_uid();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549
Cedric Le Goatera03fcb72006-10-02 02:17:26 -0700550 kill_pid_info(info->notify.sigev_signo,
551 &sig_i, info->notify_owner);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552 break;
553 case SIGEV_THREAD:
554 set_cookie(info->notify_cookie, NOTIFY_WOKENUP);
Denis V. Lunev7ee015e2007-10-10 21:14:03 -0700555 netlink_sendskb(info->notify_sock, info->notify_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700556 break;
557 }
558 /* after notification unregisters process */
Cedric Le Goatera03fcb72006-10-02 02:17:26 -0700559 put_pid(info->notify_owner);
560 info->notify_owner = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561 }
562 wake_up(&info->wait_q);
563}
564
Carsten Emde9ca7d8e2010-04-02 22:40:20 +0200565static int prepare_timeout(const struct timespec __user *u_abs_timeout,
566 ktime_t *expires, struct timespec *ts)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567{
Carsten Emde9ca7d8e2010-04-02 22:40:20 +0200568 if (copy_from_user(ts, u_abs_timeout, sizeof(struct timespec)))
569 return -EFAULT;
570 if (!timespec_valid(ts))
571 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572
Carsten Emde9ca7d8e2010-04-02 22:40:20 +0200573 *expires = timespec_to_ktime(*ts);
574 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575}
576
577static void remove_notification(struct mqueue_inode_info *info)
578{
Cedric Le Goatera03fcb72006-10-02 02:17:26 -0700579 if (info->notify_owner != NULL &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700580 info->notify.sigev_notify == SIGEV_THREAD) {
581 set_cookie(info->notify_cookie, NOTIFY_REMOVED);
Denis V. Lunev7ee015e2007-10-10 21:14:03 -0700582 netlink_sendskb(info->notify_sock, info->notify_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583 }
Cedric Le Goatera03fcb72006-10-02 02:17:26 -0700584 put_pid(info->notify_owner);
585 info->notify_owner = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700586}
587
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700588static int mq_attr_ok(struct ipc_namespace *ipc_ns, struct mq_attr *attr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700589{
590 if (attr->mq_maxmsg <= 0 || attr->mq_msgsize <= 0)
591 return 0;
592 if (capable(CAP_SYS_RESOURCE)) {
593 if (attr->mq_maxmsg > HARD_MSGMAX)
594 return 0;
595 } else {
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700596 if (attr->mq_maxmsg > ipc_ns->mq_msg_max ||
597 attr->mq_msgsize > ipc_ns->mq_msgsize_max)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598 return 0;
599 }
600 /* check for overflow */
601 if (attr->mq_msgsize > ULONG_MAX/attr->mq_maxmsg)
602 return 0;
André Goddard Rosa8834cf72010-02-23 04:04:24 -0300603 if ((unsigned long)(attr->mq_maxmsg * (attr->mq_msgsize
604 + sizeof (struct msg_msg *))) <
Linus Torvalds1da177e2005-04-16 15:20:36 -0700605 (unsigned long)(attr->mq_maxmsg * attr->mq_msgsize))
606 return 0;
607 return 1;
608}
609
610/*
611 * Invoked when creating a new queue via sys_mq_open
612 */
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700613static struct file *do_create(struct ipc_namespace *ipc_ns, struct dentry *dir,
614 struct dentry *dentry, int oflag, mode_t mode,
615 struct mq_attr *attr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616{
David Howells745ca242008-11-14 10:39:22 +1100617 const struct cred *cred = current_cred();
Dave Hansen4a3fd212008-02-15 14:37:48 -0800618 struct file *result;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619 int ret;
620
Al Viro564f6992008-12-14 04:02:26 -0500621 if (attr) {
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300622 if (!mq_attr_ok(ipc_ns, attr)) {
623 ret = -EINVAL;
Alexander Viro7c7dce92006-01-14 15:29:55 -0500624 goto out;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300625 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700626 /* store for use during create */
Al Viro564f6992008-12-14 04:02:26 -0500627 dentry->d_fsdata = attr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700628 }
629
Al Viroce3b0f82009-03-29 19:08:22 -0400630 mode &= ~current_umask();
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700631 ret = mnt_want_write(ipc_ns->mq_mnt);
Dave Hansen4a3fd212008-02-15 14:37:48 -0800632 if (ret)
633 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634 ret = vfs_create(dir->d_inode, dentry, mode, NULL);
635 dentry->d_fsdata = NULL;
636 if (ret)
Dave Hansen4a3fd212008-02-15 14:37:48 -0800637 goto out_drop_write;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700638
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700639 result = dentry_open(dentry, ipc_ns->mq_mnt, oflag, cred);
Dave Hansen4a3fd212008-02-15 14:37:48 -0800640 /*
641 * dentry_open() took a persistent mnt_want_write(),
642 * so we can now drop this one.
643 */
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700644 mnt_drop_write(ipc_ns->mq_mnt);
Dave Hansen4a3fd212008-02-15 14:37:48 -0800645 return result;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700646
Dave Hansen4a3fd212008-02-15 14:37:48 -0800647out_drop_write:
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700648 mnt_drop_write(ipc_ns->mq_mnt);
Alexander Viro7c7dce92006-01-14 15:29:55 -0500649out:
650 dput(dentry);
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700651 mntput(ipc_ns->mq_mnt);
Alexander Viro7c7dce92006-01-14 15:29:55 -0500652 return ERR_PTR(ret);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653}
654
655/* Opens existing queue */
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700656static struct file *do_open(struct ipc_namespace *ipc_ns,
657 struct dentry *dentry, int oflag)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658{
André Goddard Rosa04db0dd2010-02-23 04:04:25 -0300659 int ret;
David Howells745ca242008-11-14 10:39:22 +1100660 const struct cred *cred = current_cred();
661
662 static const int oflag2acc[O_ACCMODE] = { MAY_READ, MAY_WRITE,
663 MAY_READ | MAY_WRITE };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664
Alexander Viro7c7dce92006-01-14 15:29:55 -0500665 if ((oflag & O_ACCMODE) == (O_RDWR | O_WRONLY)) {
André Goddard Rosa04db0dd2010-02-23 04:04:25 -0300666 ret = -EINVAL;
667 goto err;
Alexander Viro7c7dce92006-01-14 15:29:55 -0500668 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669
Al Virof419a2e2008-07-22 00:07:17 -0400670 if (inode_permission(dentry->d_inode, oflag2acc[oflag & O_ACCMODE])) {
André Goddard Rosa04db0dd2010-02-23 04:04:25 -0300671 ret = -EACCES;
672 goto err;
Alexander Viro7c7dce92006-01-14 15:29:55 -0500673 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700675 return dentry_open(dentry, ipc_ns->mq_mnt, oflag, cred);
André Goddard Rosa04db0dd2010-02-23 04:04:25 -0300676
677err:
678 dput(dentry);
679 mntput(ipc_ns->mq_mnt);
680 return ERR_PTR(ret);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681}
682
Heiko Carstensd5460c92009-01-14 14:14:27 +0100683SYSCALL_DEFINE4(mq_open, const char __user *, u_name, int, oflag, mode_t, mode,
684 struct mq_attr __user *, u_attr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700685{
686 struct dentry *dentry;
687 struct file *filp;
688 char *name;
Al Viro564f6992008-12-14 04:02:26 -0500689 struct mq_attr attr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700690 int fd, error;
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700691 struct ipc_namespace *ipc_ns = current->nsproxy->ipc_ns;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692
Al Viro564f6992008-12-14 04:02:26 -0500693 if (u_attr && copy_from_user(&attr, u_attr, sizeof(struct mq_attr)))
694 return -EFAULT;
695
696 audit_mq_open(oflag, mode, u_attr ? &attr : NULL);
George C. Wilson20ca73b2006-05-24 16:09:55 -0500697
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698 if (IS_ERR(name = getname(u_name)))
699 return PTR_ERR(name);
700
Ulrich Drepper269f2132008-05-03 15:28:45 -0400701 fd = get_unused_fd_flags(O_CLOEXEC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700702 if (fd < 0)
703 goto out_putname;
704
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700705 mutex_lock(&ipc_ns->mq_mnt->mnt_root->d_inode->i_mutex);
706 dentry = lookup_one_len(name, ipc_ns->mq_mnt->mnt_root, strlen(name));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700707 if (IS_ERR(dentry)) {
708 error = PTR_ERR(dentry);
André Goddard Rosa4294a8e2010-02-23 04:04:28 -0300709 goto out_putfd;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710 }
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700711 mntget(ipc_ns->mq_mnt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700712
713 if (oflag & O_CREAT) {
714 if (dentry->d_inode) { /* entry already exists */
Al Viro5a190ae2007-06-07 12:19:32 -0400715 audit_inode(name, dentry);
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300716 if (oflag & O_EXCL) {
717 error = -EEXIST;
Alexander Viro7c7dce92006-01-14 15:29:55 -0500718 goto out;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300719 }
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700720 filp = do_open(ipc_ns, dentry, oflag);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700721 } else {
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700722 filp = do_create(ipc_ns, ipc_ns->mq_mnt->mnt_root,
723 dentry, oflag, mode,
Al Viro564f6992008-12-14 04:02:26 -0500724 u_attr ? &attr : NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725 }
Alexander Viro7c7dce92006-01-14 15:29:55 -0500726 } else {
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300727 if (!dentry->d_inode) {
728 error = -ENOENT;
Alexander Viro7c7dce92006-01-14 15:29:55 -0500729 goto out;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300730 }
Al Viro5a190ae2007-06-07 12:19:32 -0400731 audit_inode(name, dentry);
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700732 filp = do_open(ipc_ns, dentry, oflag);
Alexander Viro7c7dce92006-01-14 15:29:55 -0500733 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700734
735 if (IS_ERR(filp)) {
736 error = PTR_ERR(filp);
737 goto out_putfd;
738 }
739
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740 fd_install(fd, filp);
741 goto out_upsem;
742
Alexander Viro7c7dce92006-01-14 15:29:55 -0500743out:
744 dput(dentry);
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700745 mntput(ipc_ns->mq_mnt);
Alexander Viro7c7dce92006-01-14 15:29:55 -0500746out_putfd:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747 put_unused_fd(fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748 fd = error;
749out_upsem:
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700750 mutex_unlock(&ipc_ns->mq_mnt->mnt_root->d_inode->i_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700751out_putname:
752 putname(name);
753 return fd;
754}
755
Heiko Carstensd5460c92009-01-14 14:14:27 +0100756SYSCALL_DEFINE1(mq_unlink, const char __user *, u_name)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700757{
758 int err;
759 char *name;
760 struct dentry *dentry;
761 struct inode *inode = NULL;
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -0700762 struct ipc_namespace *ipc_ns = current->nsproxy->ipc_ns;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763
764 name = getname(u_name);
765 if (IS_ERR(name))
766 return PTR_ERR(name);
767
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700768 mutex_lock_nested(&ipc_ns->mq_mnt->mnt_root->d_inode->i_mutex,
Peter Zijlstra7a434812007-03-06 01:42:09 -0800769 I_MUTEX_PARENT);
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700770 dentry = lookup_one_len(name, ipc_ns->mq_mnt->mnt_root, strlen(name));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700771 if (IS_ERR(dentry)) {
772 err = PTR_ERR(dentry);
773 goto out_unlock;
774 }
775
776 if (!dentry->d_inode) {
777 err = -ENOENT;
778 goto out_err;
779 }
780
781 inode = dentry->d_inode;
782 if (inode)
Al Viro7de9c6e2010-10-23 11:11:40 -0400783 ihold(inode);
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700784 err = mnt_want_write(ipc_ns->mq_mnt);
Dave Hansen06227532008-02-15 14:37:34 -0800785 if (err)
786 goto out_err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787 err = vfs_unlink(dentry->d_parent->d_inode, dentry);
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700788 mnt_drop_write(ipc_ns->mq_mnt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789out_err:
790 dput(dentry);
791
792out_unlock:
Serge E. Hallyn614b84c2009-04-06 19:01:08 -0700793 mutex_unlock(&ipc_ns->mq_mnt->mnt_root->d_inode->i_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794 putname(name);
795 if (inode)
796 iput(inode);
797
798 return err;
799}
800
801/* Pipelined send and receive functions.
802 *
803 * If a receiver finds no waiting message, then it registers itself in the
804 * list of waiting receivers. A sender checks that list before adding the new
805 * message into the message array. If there is a waiting receiver, then it
806 * bypasses the message array and directly hands the message over to the
807 * receiver.
808 * The receiver accepts the message and returns without grabbing the queue
809 * spinlock. Therefore an intermediate STATE_PENDING state and memory barriers
810 * are necessary. The same algorithm is used for sysv semaphores, see
Serge E. Hallyn7b7a3172006-03-28 01:56:23 -0800811 * ipc/sem.c for more details.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700812 *
813 * The same algorithm is used for senders.
814 */
815
816/* pipelined_send() - send a message directly to the task waiting in
817 * sys_mq_timedreceive() (without inserting message into a queue).
818 */
819static inline void pipelined_send(struct mqueue_inode_info *info,
820 struct msg_msg *message,
821 struct ext_wait_queue *receiver)
822{
823 receiver->msg = message;
824 list_del(&receiver->list);
825 receiver->state = STATE_PENDING;
826 wake_up_process(receiver->task);
akpm@osdl.orgd59dd462005-05-01 08:58:47 -0700827 smp_wmb();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828 receiver->state = STATE_READY;
829}
830
831/* pipelined_receive() - if there is task waiting in sys_mq_timedsend()
832 * gets its message and put to the queue (we have one free place for sure). */
833static inline void pipelined_receive(struct mqueue_inode_info *info)
834{
835 struct ext_wait_queue *sender = wq_get_first_waiter(info, SEND);
836
837 if (!sender) {
838 /* for poll */
839 wake_up_interruptible(&info->wait_q);
840 return;
841 }
842 msg_insert(sender->msg, info);
843 list_del(&sender->list);
844 sender->state = STATE_PENDING;
845 wake_up_process(sender->task);
akpm@osdl.orgd59dd462005-05-01 08:58:47 -0700846 smp_wmb();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700847 sender->state = STATE_READY;
848}
849
Heiko Carstensc4ea37c2009-01-14 14:14:28 +0100850SYSCALL_DEFINE5(mq_timedsend, mqd_t, mqdes, const char __user *, u_msg_ptr,
851 size_t, msg_len, unsigned int, msg_prio,
852 const struct timespec __user *, u_abs_timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853{
854 struct file *filp;
855 struct inode *inode;
856 struct ext_wait_queue wait;
857 struct ext_wait_queue *receiver;
858 struct msg_msg *msg_ptr;
859 struct mqueue_inode_info *info;
Carsten Emde9ca7d8e2010-04-02 22:40:20 +0200860 ktime_t expires, *timeout = NULL;
861 struct timespec ts;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700862 int ret;
863
Al Viroc32c8af2008-12-14 03:46:48 -0500864 if (u_abs_timeout) {
Carsten Emde9ca7d8e2010-04-02 22:40:20 +0200865 int res = prepare_timeout(u_abs_timeout, &expires, &ts);
866 if (res)
867 return res;
868 timeout = &expires;
Al Viroc32c8af2008-12-14 03:46:48 -0500869 }
George C. Wilson20ca73b2006-05-24 16:09:55 -0500870
Linus Torvalds1da177e2005-04-16 15:20:36 -0700871 if (unlikely(msg_prio >= (unsigned long) MQ_PRIO_MAX))
872 return -EINVAL;
873
Carsten Emde9ca7d8e2010-04-02 22:40:20 +0200874 audit_mq_sendrecv(mqdes, msg_len, msg_prio, timeout ? &ts : NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700875
Linus Torvalds1da177e2005-04-16 15:20:36 -0700876 filp = fget(mqdes);
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300877 if (unlikely(!filp)) {
878 ret = -EBADF;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879 goto out;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300880 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700881
Josef Sipek6d630792006-12-08 02:37:11 -0800882 inode = filp->f_path.dentry->d_inode;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300883 if (unlikely(filp->f_op != &mqueue_file_operations)) {
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 info = MQUEUE_I(inode);
Al Viro5a190ae2007-06-07 12:19:32 -0400888 audit_inode(NULL, filp->f_path.dentry);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300890 if (unlikely(!(filp->f_mode & FMODE_WRITE))) {
891 ret = -EBADF;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700892 goto out_fput;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300893 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700894
895 if (unlikely(msg_len > info->attr.mq_msgsize)) {
896 ret = -EMSGSIZE;
897 goto out_fput;
898 }
899
900 /* First try to allocate memory, before doing anything with
901 * existing queues. */
902 msg_ptr = load_msg(u_msg_ptr, msg_len);
903 if (IS_ERR(msg_ptr)) {
904 ret = PTR_ERR(msg_ptr);
905 goto out_fput;
906 }
907 msg_ptr->m_ts = msg_len;
908 msg_ptr->m_type = msg_prio;
909
910 spin_lock(&info->lock);
911
912 if (info->attr.mq_curmsgs == info->attr.mq_maxmsg) {
913 if (filp->f_flags & O_NONBLOCK) {
914 spin_unlock(&info->lock);
915 ret = -EAGAIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916 } else {
917 wait.task = current;
918 wait.msg = (void *) msg_ptr;
919 wait.state = STATE_NONE;
920 ret = wq_sleep(info, SEND, timeout, &wait);
921 }
922 if (ret < 0)
923 free_msg(msg_ptr);
924 } else {
925 receiver = wq_get_first_waiter(info, RECV);
926 if (receiver) {
927 pipelined_send(info, msg_ptr, receiver);
928 } else {
929 /* adds message to the queue */
930 msg_insert(msg_ptr, info);
931 __do_notify(info);
932 }
933 inode->i_atime = inode->i_mtime = inode->i_ctime =
934 CURRENT_TIME;
935 spin_unlock(&info->lock);
936 ret = 0;
937 }
938out_fput:
939 fput(filp);
940out:
941 return ret;
942}
943
Heiko Carstensc4ea37c2009-01-14 14:14:28 +0100944SYSCALL_DEFINE5(mq_timedreceive, mqd_t, mqdes, char __user *, u_msg_ptr,
945 size_t, msg_len, unsigned int __user *, u_msg_prio,
946 const struct timespec __user *, u_abs_timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700948 ssize_t ret;
949 struct msg_msg *msg_ptr;
950 struct file *filp;
951 struct inode *inode;
952 struct mqueue_inode_info *info;
953 struct ext_wait_queue wait;
Carsten Emde9ca7d8e2010-04-02 22:40:20 +0200954 ktime_t expires, *timeout = NULL;
955 struct timespec ts;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956
Al Viroc32c8af2008-12-14 03:46:48 -0500957 if (u_abs_timeout) {
Carsten Emde9ca7d8e2010-04-02 22:40:20 +0200958 int res = prepare_timeout(u_abs_timeout, &expires, &ts);
959 if (res)
960 return res;
961 timeout = &expires;
Al Viroc32c8af2008-12-14 03:46:48 -0500962 }
George C. Wilson20ca73b2006-05-24 16:09:55 -0500963
Carsten Emde9ca7d8e2010-04-02 22:40:20 +0200964 audit_mq_sendrecv(mqdes, msg_len, 0, timeout ? &ts : NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965
Linus Torvalds1da177e2005-04-16 15:20:36 -0700966 filp = fget(mqdes);
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300967 if (unlikely(!filp)) {
968 ret = -EBADF;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700969 goto out;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300970 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700971
Josef Sipek6d630792006-12-08 02:37:11 -0800972 inode = filp->f_path.dentry->d_inode;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300973 if (unlikely(filp->f_op != &mqueue_file_operations)) {
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 info = MQUEUE_I(inode);
Al Viro5a190ae2007-06-07 12:19:32 -0400978 audit_inode(NULL, filp->f_path.dentry);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700979
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300980 if (unlikely(!(filp->f_mode & FMODE_READ))) {
981 ret = -EBADF;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700982 goto out_fput;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -0300983 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700984
985 /* checks if buffer is big enough */
986 if (unlikely(msg_len < info->attr.mq_msgsize)) {
987 ret = -EMSGSIZE;
988 goto out_fput;
989 }
990
991 spin_lock(&info->lock);
992 if (info->attr.mq_curmsgs == 0) {
993 if (filp->f_flags & O_NONBLOCK) {
994 spin_unlock(&info->lock);
995 ret = -EAGAIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700996 } else {
997 wait.task = current;
998 wait.state = STATE_NONE;
999 ret = wq_sleep(info, RECV, timeout, &wait);
1000 msg_ptr = wait.msg;
1001 }
1002 } else {
1003 msg_ptr = msg_get(info);
1004
1005 inode->i_atime = inode->i_mtime = inode->i_ctime =
1006 CURRENT_TIME;
1007
1008 /* There is now free space in queue. */
1009 pipelined_receive(info);
1010 spin_unlock(&info->lock);
1011 ret = 0;
1012 }
1013 if (ret == 0) {
1014 ret = msg_ptr->m_ts;
1015
1016 if ((u_msg_prio && put_user(msg_ptr->m_type, u_msg_prio)) ||
1017 store_msg(u_msg_ptr, msg_ptr, msg_ptr->m_ts)) {
1018 ret = -EFAULT;
1019 }
1020 free_msg(msg_ptr);
1021 }
1022out_fput:
1023 fput(filp);
1024out:
1025 return ret;
1026}
1027
1028/*
1029 * Notes: the case when user wants us to deregister (with NULL as pointer)
1030 * and he isn't currently owner of notification, will be silently discarded.
1031 * It isn't explicitly defined in the POSIX.
1032 */
Heiko Carstensc4ea37c2009-01-14 14:14:28 +01001033SYSCALL_DEFINE2(mq_notify, mqd_t, mqdes,
1034 const struct sigevent __user *, u_notification)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001035{
1036 int ret;
1037 struct file *filp;
1038 struct sock *sock;
1039 struct inode *inode;
1040 struct sigevent notification;
1041 struct mqueue_inode_info *info;
1042 struct sk_buff *nc;
1043
Al Viro20114f72008-12-10 07:16:12 -05001044 if (u_notification) {
1045 if (copy_from_user(&notification, u_notification,
1046 sizeof(struct sigevent)))
1047 return -EFAULT;
1048 }
1049
1050 audit_mq_notify(mqdes, u_notification ? &notification : NULL);
George C. Wilson20ca73b2006-05-24 16:09:55 -05001051
Linus Torvalds1da177e2005-04-16 15:20:36 -07001052 nc = NULL;
1053 sock = NULL;
1054 if (u_notification != NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001055 if (unlikely(notification.sigev_notify != SIGEV_NONE &&
1056 notification.sigev_notify != SIGEV_SIGNAL &&
1057 notification.sigev_notify != SIGEV_THREAD))
1058 return -EINVAL;
1059 if (notification.sigev_notify == SIGEV_SIGNAL &&
Jesper Juhl7ed20e12005-05-01 08:59:14 -07001060 !valid_signal(notification.sigev_signo)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001061 return -EINVAL;
1062 }
1063 if (notification.sigev_notify == SIGEV_THREAD) {
Patrick McHardyc3d8d1e2007-11-07 02:42:09 -08001064 long timeo;
1065
Linus Torvalds1da177e2005-04-16 15:20:36 -07001066 /* create the notify skb */
1067 nc = alloc_skb(NOTIFY_COOKIE_LEN, GFP_KERNEL);
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -03001068 if (!nc) {
1069 ret = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001070 goto out;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -03001071 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001072 if (copy_from_user(nc->data,
1073 notification.sigev_value.sival_ptr,
1074 NOTIFY_COOKIE_LEN)) {
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -03001075 ret = -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001076 goto out;
1077 }
1078
1079 /* TODO: add a header? */
1080 skb_put(nc, NOTIFY_COOKIE_LEN);
1081 /* and attach it to the socket */
1082retry:
1083 filp = fget(notification.sigev_signo);
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -03001084 if (!filp) {
1085 ret = -EBADF;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001086 goto out;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -03001087 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001088 sock = netlink_getsockbyfilp(filp);
1089 fput(filp);
1090 if (IS_ERR(sock)) {
1091 ret = PTR_ERR(sock);
1092 sock = NULL;
1093 goto out;
1094 }
1095
Patrick McHardyc3d8d1e2007-11-07 02:42:09 -08001096 timeo = MAX_SCHEDULE_TIMEOUT;
Denis V. Lunev9457afe2008-06-05 11:23:39 -07001097 ret = netlink_attachskb(sock, nc, &timeo, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001098 if (ret == 1)
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -03001099 goto retry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001100 if (ret) {
1101 sock = NULL;
1102 nc = NULL;
1103 goto out;
1104 }
1105 }
1106 }
1107
Linus Torvalds1da177e2005-04-16 15:20:36 -07001108 filp = fget(mqdes);
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -03001109 if (!filp) {
1110 ret = -EBADF;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001111 goto out;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -03001112 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113
Josef Sipek6d630792006-12-08 02:37:11 -08001114 inode = filp->f_path.dentry->d_inode;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -03001115 if (unlikely(filp->f_op != &mqueue_file_operations)) {
1116 ret = -EBADF;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001117 goto out_fput;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -03001118 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001119 info = MQUEUE_I(inode);
1120
1121 ret = 0;
1122 spin_lock(&info->lock);
1123 if (u_notification == NULL) {
Cedric Le Goatera03fcb72006-10-02 02:17:26 -07001124 if (info->notify_owner == task_tgid(current)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125 remove_notification(info);
1126 inode->i_atime = inode->i_ctime = CURRENT_TIME;
1127 }
Cedric Le Goatera03fcb72006-10-02 02:17:26 -07001128 } else if (info->notify_owner != NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129 ret = -EBUSY;
1130 } else {
1131 switch (notification.sigev_notify) {
1132 case SIGEV_NONE:
1133 info->notify.sigev_notify = SIGEV_NONE;
1134 break;
1135 case SIGEV_THREAD:
1136 info->notify_sock = sock;
1137 info->notify_cookie = nc;
1138 sock = NULL;
1139 nc = NULL;
1140 info->notify.sigev_notify = SIGEV_THREAD;
1141 break;
1142 case SIGEV_SIGNAL:
1143 info->notify.sigev_signo = notification.sigev_signo;
1144 info->notify.sigev_value = notification.sigev_value;
1145 info->notify.sigev_notify = SIGEV_SIGNAL;
1146 break;
1147 }
Cedric Le Goatera03fcb72006-10-02 02:17:26 -07001148
1149 info->notify_owner = get_pid(task_tgid(current));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001150 inode->i_atime = inode->i_ctime = CURRENT_TIME;
1151 }
1152 spin_unlock(&info->lock);
1153out_fput:
1154 fput(filp);
1155out:
1156 if (sock) {
1157 netlink_detachskb(sock, nc);
1158 } else if (nc) {
1159 dev_kfree_skb(nc);
1160 }
1161 return ret;
1162}
1163
Heiko Carstensc4ea37c2009-01-14 14:14:28 +01001164SYSCALL_DEFINE3(mq_getsetattr, mqd_t, mqdes,
1165 const struct mq_attr __user *, u_mqstat,
1166 struct mq_attr __user *, u_omqstat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001167{
1168 int ret;
1169 struct mq_attr mqstat, omqstat;
1170 struct file *filp;
1171 struct inode *inode;
1172 struct mqueue_inode_info *info;
1173
1174 if (u_mqstat != NULL) {
1175 if (copy_from_user(&mqstat, u_mqstat, sizeof(struct mq_attr)))
1176 return -EFAULT;
1177 if (mqstat.mq_flags & (~O_NONBLOCK))
1178 return -EINVAL;
1179 }
1180
Linus Torvalds1da177e2005-04-16 15:20:36 -07001181 filp = fget(mqdes);
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -03001182 if (!filp) {
1183 ret = -EBADF;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184 goto out;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -03001185 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186
Josef Sipek6d630792006-12-08 02:37:11 -08001187 inode = filp->f_path.dentry->d_inode;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -03001188 if (unlikely(filp->f_op != &mqueue_file_operations)) {
1189 ret = -EBADF;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001190 goto out_fput;
André Goddard Rosa8d8ffef2010-02-23 04:04:26 -03001191 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001192 info = MQUEUE_I(inode);
1193
1194 spin_lock(&info->lock);
1195
1196 omqstat = info->attr;
1197 omqstat.mq_flags = filp->f_flags & O_NONBLOCK;
1198 if (u_mqstat) {
Al Viro73929062008-12-10 06:58:59 -05001199 audit_mq_getsetattr(mqdes, &mqstat);
Jonathan Corbetdb1dd4d2009-02-06 15:25:24 -07001200 spin_lock(&filp->f_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001201 if (mqstat.mq_flags & O_NONBLOCK)
1202 filp->f_flags |= O_NONBLOCK;
1203 else
1204 filp->f_flags &= ~O_NONBLOCK;
Jonathan Corbetdb1dd4d2009-02-06 15:25:24 -07001205 spin_unlock(&filp->f_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001206
1207 inode->i_atime = inode->i_ctime = CURRENT_TIME;
1208 }
1209
1210 spin_unlock(&info->lock);
1211
1212 ret = 0;
1213 if (u_omqstat != NULL && copy_to_user(u_omqstat, &omqstat,
1214 sizeof(struct mq_attr)))
1215 ret = -EFAULT;
1216
1217out_fput:
1218 fput(filp);
1219out:
1220 return ret;
1221}
1222
Arjan van de Ven92e1d5b2007-02-12 00:55:39 -08001223static const struct inode_operations mqueue_dir_inode_operations = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001224 .lookup = simple_lookup,
1225 .create = mqueue_create,
1226 .unlink = mqueue_unlink,
1227};
1228
Arjan van de Ven9a321442007-02-12 00:55:35 -08001229static const struct file_operations mqueue_file_operations = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001230 .flush = mqueue_flush_file,
1231 .poll = mqueue_poll_file,
1232 .read = mqueue_read_file,
Arnd Bergmann6038f372010-08-15 18:52:59 +02001233 .llseek = default_llseek,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001234};
1235
Alexey Dobriyanb87221d2009-09-21 17:01:09 -07001236static const struct super_operations mqueue_super_ops = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001237 .alloc_inode = mqueue_alloc_inode,
1238 .destroy_inode = mqueue_destroy_inode,
Al Viro6d8af642010-06-05 16:29:45 -04001239 .evict_inode = mqueue_evict_inode,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001240 .statfs = simple_statfs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001241};
1242
1243static struct file_system_type mqueue_fs_type = {
1244 .name = "mqueue",
Al Viroceefda62010-07-26 13:16:50 +04001245 .mount = mqueue_mount,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001246 .kill_sb = kill_litter_super,
1247};
1248
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -07001249int mq_init_ns(struct ipc_namespace *ns)
1250{
1251 ns->mq_queues_count = 0;
1252 ns->mq_queues_max = DFLT_QUEUESMAX;
1253 ns->mq_msg_max = DFLT_MSGMAX;
1254 ns->mq_msgsize_max = DFLT_MSGSIZEMAX;
1255
1256 ns->mq_mnt = kern_mount_data(&mqueue_fs_type, ns);
1257 if (IS_ERR(ns->mq_mnt)) {
1258 int err = PTR_ERR(ns->mq_mnt);
1259 ns->mq_mnt = NULL;
1260 return err;
1261 }
1262 return 0;
1263}
1264
1265void mq_clear_sbinfo(struct ipc_namespace *ns)
1266{
1267 ns->mq_mnt->mnt_sb->s_fs_info = NULL;
1268}
1269
1270void mq_put_mnt(struct ipc_namespace *ns)
1271{
1272 mntput(ns->mq_mnt);
1273}
1274
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275static int __init init_mqueue_fs(void)
1276{
1277 int error;
1278
1279 mqueue_inode_cachep = kmem_cache_create("mqueue_inode_cache",
1280 sizeof(struct mqueue_inode_info), 0,
Paul Mundt20c2df82007-07-20 10:11:58 +09001281 SLAB_HWCACHE_ALIGN, init_once);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001282 if (mqueue_inode_cachep == NULL)
1283 return -ENOMEM;
1284
André Goddard Rosa2329e392010-02-23 04:04:27 -03001285 /* ignore failures - they are not fatal */
Serge E. Hallynbdc8e5f2009-04-06 19:01:11 -07001286 mq_sysctl_table = mq_register_sysctl_table();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287
1288 error = register_filesystem(&mqueue_fs_type);
1289 if (error)
1290 goto out_sysctl;
1291
Serge E. Hallyn7eafd7c2009-04-06 19:01:10 -07001292 spin_lock_init(&mq_lock);
1293
1294 init_ipc_ns.mq_mnt = kern_mount_data(&mqueue_fs_type, &init_ipc_ns);
Serge E. Hallyn614b84c2009-04-06 19:01:08 -07001295 if (IS_ERR(init_ipc_ns.mq_mnt)) {
1296 error = PTR_ERR(init_ipc_ns.mq_mnt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001297 goto out_filesystem;
1298 }
1299
Linus Torvalds1da177e2005-04-16 15:20:36 -07001300 return 0;
1301
1302out_filesystem:
1303 unregister_filesystem(&mqueue_fs_type);
1304out_sysctl:
1305 if (mq_sysctl_table)
1306 unregister_sysctl_table(mq_sysctl_table);
Alexey Dobriyan1a1d92c2006-09-27 01:49:40 -07001307 kmem_cache_destroy(mqueue_inode_cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001308 return error;
1309}
1310
1311__initcall(init_mqueue_fs);