ipc/msg.c

   1 /*
   2  * linux/ipc/msg.c
   3  * Copyright (C) 1992 Krishna Balasubramanian
   4  *
   5  * Removed all the remaining kerneld mess
   6  * Catch the -EFAULT stuff properly
   7  * Use GFP_KERNEL for messages as in 1.2
   8  * Fixed up the unchecked user space derefs
   9  * Copyright (C) 1998 Alan Cox & Andi Kleen
  10  *
  11  * /proc/sysvipc/msg support (c) 1999 Dragos Acostachioaie <dragos@iname.com>
  12  *
  13  * mostly rewritten, threaded and wake-one semantics added
  14  * MSGMAX limit removed, sysctl's added
  15  * (c) 1999 Manfred Spraul <manfred@colorfullife.com>
  16  *
  17  * support for audit of ipc object properties and permission changes
  18  * Dustin Kirkland <dustin.kirkland@us.ibm.com>
  19  *
  20  * namespaces support
  21  * OpenVZ, SWsoft Inc.
  22  * Pavel Emelianov <xemul@openvz.org>
  23  */
  24
  25 #include <linux/capability.h>
  26 #include <linux/slab.h>
  27 #include <linux/msg.h>
  28 #include <linux/spinlock.h>
  29 #include <linux/init.h>
  30 #include <linux/proc_fs.h>
  31 #include <linux/list.h>
  32 #include <linux/security.h>
  33 #include <linux/sched.h>
  34 #include <linux/syscalls.h>
  35 #include <linux/audit.h>
  36 #include <linux/seq_file.h>
  37 #include <linux/rwsem.h>
  38 #include <linux/nsproxy.h>
  39 #include <linux/ipc_namespace.h>
  40
  41 #include <asm/current.h>
  42 #include <asm/uaccess.h>
  43 #include "util.h"
  44
  45 /*
  46  * one msg_receiver structure for each sleeping receiver:
  47  */
  48 struct msg_receiver {
  49         struct list_head        r_list;
  50         struct task_struct      *r_tsk;
  51
  52         int                     r_mode;
  53         long                    r_msgtype;
  54         long                    r_maxsize;
  55
  56         struct msg_msg          *volatile r_msg;
  57 };
  58
  59 /* one msg_sender for each sleeping sender */
  60 struct msg_sender {
  61         struct list_head        list;
  62         struct task_struct      *tsk;
  63 };
  64
  65 #define SEARCH_ANY              1
  66 #define SEARCH_EQUAL            2
  67 #define SEARCH_NOTEQUAL         3
  68 #define SEARCH_LESSEQUAL        4
  69
  70 #define msg_ids(ns)     ((ns)->ids[IPC_MSG_IDS])
  71
  72 #define msg_unlock(msq)         ipc_unlock(&(msq)->q_perm)
  73 #define msg_buildid(id, seq)    ipc_buildid(id, seq)
  74
  75 static void freeque(struct ipc_namespace *, struct msg_queue *);
  76 static int newque(struct ipc_namespace *, struct ipc_params *);
  77 #ifdef CONFIG_PROC_FS
  78 static int sysvipc_msg_proc_show(struct seq_file *s, void *it);
  79 #endif
  80
  81 void msg_init_ns(struct ipc_namespace *ns)
  82 {
  83         ns->msg_ctlmax = MSGMAX;
  84         ns->msg_ctlmnb = MSGMNB;
  85         ns->msg_ctlmni = MSGMNI;
  86         atomic_set(&ns->msg_bytes, 0);
  87         atomic_set(&ns->msg_hdrs, 0);
  88         ipc_init_ids(&ns->ids[IPC_MSG_IDS]);
  89 }
  90
  91 #ifdef CONFIG_IPC_NS
  92 void msg_exit_ns(struct ipc_namespace *ns)
  93 {
  94         struct msg_queue *msq;
  95         struct kern_ipc_perm *perm;
  96         int next_id;
  97         int total, in_use;
  98
  99         down_write(&msg_ids(ns).rw_mutex);
 100
 101         in_use = msg_ids(ns).in_use;
 102
 103         for (total = 0, next_id = 0; total < in_use; next_id++) {
 104                 perm = idr_find(&msg_ids(ns).ipcs_idr, next_id);
 105                 if (perm == NULL)
 106                         continue;
 107                 ipc_lock_by_ptr(perm);
 108                 msq = container_of(perm, struct msg_queue, q_perm);
 109                 freeque(ns, msq);
 110                 total++;
 111         }
 112
 113         up_write(&msg_ids(ns).rw_mutex);
 114 }
 115 #endif
 116
 117 void __init msg_init(void)
 118 {
 119         msg_init_ns(&init_ipc_ns);
 120         ipc_init_proc_interface("sysvipc/msg",
 121                                 "       key      msqid perms      cbytes       qnum lspid lrpid   uid   gid  cuid  cgid      stime      rtime      ctime\n",
 122                                 IPC_MSG_IDS, sysvipc_msg_proc_show);
 123 }
 124
 125 /*
 126  * This routine is called in the paths where the rw_mutex is held to protect
 127  * access to the idr tree.
 128  */
 129 static inline struct msg_queue *msg_lock_check_down(struct ipc_namespace *ns,
 130                                                 int id)
 131 {
 132         struct kern_ipc_perm *ipcp = ipc_lock_check_down(&msg_ids(ns), id);
 133
 134         if (IS_ERR(ipcp))
 135                 return (struct msg_queue *)ipcp;
 136
 137         return container_of(ipcp, struct msg_queue, q_perm);
 138 }
 139
 140 /*
 141  * msg_lock_(check_) routines are called in the paths where the rw_mutex
 142  * is not held.
 143  */
 144 static inline struct msg_queue *msg_lock(struct ipc_namespace *ns, int id)
 145 {
 146         struct kern_ipc_perm *ipcp = ipc_lock(&msg_ids(ns), id);
 147
 148         if (IS_ERR(ipcp))
 149                 return (struct msg_queue *)ipcp;
 150
 151         return container_of(ipcp, struct msg_queue, q_perm);
 152 }
 153
 154 static inline struct msg_queue *msg_lock_check(struct ipc_namespace *ns,
 155                                                 int id)
 156 {
 157         struct kern_ipc_perm *ipcp = ipc_lock_check(&msg_ids(ns), id);
 158
 159         if (IS_ERR(ipcp))
 160                 return (struct msg_queue *)ipcp;
 161
 162         return container_of(ipcp, struct msg_queue, q_perm);
 163 }
 164
 165 static inline void msg_rmid(struct ipc_namespace *ns, struct msg_queue *s)
 166 {
 167         ipc_rmid(&msg_ids(ns), &s->q_perm);
 168 }
 169
 170 /**
 171  * newque - Create a new msg queue
 172  * @ns: namespace
 173  * @params: ptr to the structure that contains the key and msgflg
 174  *
 175  * Called with msg_ids.rw_mutex held (writer)
 176  */
 177 static int newque(struct ipc_namespace *ns, struct ipc_params *params)
 178 {
 179         struct msg_queue *msq;
 180         int id, retval;
 181         key_t key = params->key;
 182         int msgflg = params->flg;
 183
 184         msq = ipc_rcu_alloc(sizeof(*msq));
 185         if (!msq)
 186                 return -ENOMEM;
 187
 188         msq->q_perm.mode = msgflg & S_IRWXUGO;
 189         msq->q_perm.key = key;
 190
 191         msq->q_perm.security = NULL;
 192         retval = security_msg_queue_alloc(msq);
 193         if (retval) {
 194                 ipc_rcu_putref(msq);
 195                 return retval;
 196         }
 197
 198         /*
 199          * ipc_addid() locks msq
 200          */
 201         id = ipc_addid(&msg_ids(ns), &msq->q_perm, ns->msg_ctlmni);
 202         if (id < 0) {
 203                 security_msg_queue_free(msq);
 204                 ipc_rcu_putref(msq);
 205                 return id;
 206         }
 207
 208         msq->q_perm.id = msg_buildid(id, msq->q_perm.seq);
 209         msq->q_stime = msq->q_rtime = 0;
 210         msq->q_ctime = get_seconds();
 211         msq->q_cbytes = msq->q_qnum = 0;
 212         msq->q_qbytes = ns->msg_ctlmnb;
 213         msq->q_lspid = msq->q_lrpid = 0;
 214         INIT_LIST_HEAD(&msq->q_messages);
 215         INIT_LIST_HEAD(&msq->q_receivers);
 216         INIT_LIST_HEAD(&msq->q_senders);
 217
 218         msg_unlock(msq);
 219
 220         return msq->q_perm.id;
 221 }
 222
 223 static inline void ss_add(struct msg_queue *msq, struct msg_sender *mss)
 224 {
 225         mss->tsk = current;
 226         current->state = TASK_INTERRUPTIBLE;
 227         list_add_tail(&mss->list, &msq->q_senders);
 228 }
 229
 230 static inline void ss_del(struct msg_sender *mss)
 231 {
 232         if (mss->list.next != NULL)
 233                 list_del(&mss->list);
 234 }
 235
 236 static void ss_wakeup(struct list_head *h, int kill)
 237 {
 238         struct list_head *tmp;
 239
 240         tmp = h->next;
 241         while (tmp != h) {
 242                 struct msg_sender *mss;
 243
 244                 mss = list_entry(tmp, struct msg_sender, list);
 245                 tmp = tmp->next;
 246                 if (kill)
 247                         mss->list.next = NULL;
 248                 wake_up_process(mss->tsk);
 249         }
 250 }
 251
 252 static void expunge_all(struct msg_queue *msq, int res)
 253 {
 254         struct list_head *tmp;
 255
 256         tmp = msq->q_receivers.next;
 257         while (tmp != &msq->q_receivers) {
 258                 struct msg_receiver *msr;
 259
 260                 msr = list_entry(tmp, struct msg_receiver, r_list);
 261                 tmp = tmp->next;
 262                 msr->r_msg = NULL;
 263                 wake_up_process(msr->r_tsk);
 264                 smp_mb();
 265                 msr->r_msg = ERR_PTR(res);
 266         }
 267 }
 268
 269 /*
 270  * freeque() wakes up waiters on the sender and receiver waiting queue,
 271  * removes the message queue from message queue ID IDR, and cleans up all the
 272  * messages associated with this queue.
 273  *
 274  * msg_ids.rw_mutex (writer) and the spinlock for this message queue are held
 275  * before freeque() is called. msg_ids.rw_mutex remains locked on exit.
 276  */
 277 static void freeque(struct ipc_namespace *ns, struct msg_queue *msq)
 278 {
 279         struct list_head *tmp;
 280
 281         expunge_all(msq, -EIDRM);
 282         ss_wakeup(&msq->q_senders, 1);
 283         msg_rmid(ns, msq);
 284         msg_unlock(msq);
 285
 286         tmp = msq->q_messages.next;
 287         while (tmp != &msq->q_messages) {
 288                 struct msg_msg *msg = list_entry(tmp, struct msg_msg, m_list);
 289
 290                 tmp = tmp->next;
 291                 atomic_dec(&ns->msg_hdrs);
 292                 free_msg(msg);
 293         }
 294         atomic_sub(msq->q_cbytes, &ns->msg_bytes);
 295         security_msg_queue_free(msq);
 296         ipc_rcu_putref(msq);
 297 }
 298
 299 /*
 300  * Called with msg_ids.rw_mutex and ipcp locked.
 301  */
 302 static inline int msg_security(struct kern_ipc_perm *ipcp, int msgflg)
 303 {
 304         struct msg_queue *msq = container_of(ipcp, struct msg_queue, q_perm);
 305
 306         return security_msg_queue_associate(msq, msgflg);
 307 }
 308
 309 asmlinkage long sys_msgget(key_t key, int msgflg)
 310 {
 311         struct ipc_namespace *ns;
 312         struct ipc_ops msg_ops;
 313         struct ipc_params msg_params;
 314
 315         ns = current->nsproxy->ipc_ns;
 316
 317         msg_ops.getnew = newque;
 318         msg_ops.associate = msg_security;
 319         msg_ops.more_checks = NULL;
 320
 321         msg_params.key = key;
 322         msg_params.flg = msgflg;
 323
 324         return ipcget(ns, &msg_ids(ns), &msg_ops, &msg_params);
 325 }
 326
 327 static inline unsigned long
 328 copy_msqid_to_user(void __user *buf, struct msqid64_ds *in, int version)
 329 {
 330         switch(version) {
 331         case IPC_64:
 332                 return copy_to_user(buf, in, sizeof(*in));
 333         case IPC_OLD:
 334         {
 335                 struct msqid_ds out;
 336
 337                 memset(&out, 0, sizeof(out));
 338
 339                 ipc64_perm_to_ipc_perm(&in->msg_perm, &out.msg_perm);
 340
 341                 out.msg_stime           = in->msg_stime;
 342                 out.msg_rtime           = in->msg_rtime;
 343                 out.msg_ctime           = in->msg_ctime;
 344
 345                 if (in->msg_cbytes > USHRT_MAX)
 346                         out.msg_cbytes  = USHRT_MAX;
 347                 else
 348                         out.msg_cbytes  = in->msg_cbytes;
 349                 out.msg_lcbytes         = in->msg_cbytes;
 350
 351                 if (in->msg_qnum > USHRT_MAX)
 352                         out.msg_qnum    = USHRT_MAX;
 353                 else
 354                         out.msg_qnum    = in->msg_qnum;
 355
 356                 if (in->msg_qbytes > USHRT_MAX)
 357                         out.msg_qbytes  = USHRT_MAX;
 358                 else
 359                         out.msg_qbytes  = in->msg_qbytes;
 360                 out.msg_lqbytes         = in->msg_qbytes;
 361
 362                 out.msg_lspid           = in->msg_lspid;
 363                 out.msg_lrpid           = in->msg_lrpid;
 364
 365                 return copy_to_user(buf, &out, sizeof(out));
 366         }
 367         default:
 368                 return -EINVAL;
 369         }
 370 }
 371
 372 struct msq_setbuf {
 373         unsigned long   qbytes;
 374         uid_t           uid;
 375         gid_t           gid;
 376         mode_t          mode;
 377 };
 378
 379 static inline unsigned long
 380 copy_msqid_from_user(struct msq_setbuf *out, void __user *buf, int version)
 381 {
 382         switch(version) {
 383         case IPC_64:
 384         {
 385                 struct msqid64_ds tbuf;
 386
 387                 if (copy_from_user(&tbuf, buf, sizeof(tbuf)))
 388                         return -EFAULT;
 389
 390                 out->qbytes             = tbuf.msg_qbytes;
 391                 out->uid                = tbuf.msg_perm.uid;
 392                 out->gid                = tbuf.msg_perm.gid;
 393                 out->mode               = tbuf.msg_perm.mode;
 394
 395                 return 0;
 396         }
 397         case IPC_OLD:
 398         {
 399                 struct msqid_ds tbuf_old;
 400
 401                 if (copy_from_user(&tbuf_old, buf, sizeof(tbuf_old)))
 402                         return -EFAULT;
 403
 404                 out->uid                = tbuf_old.msg_perm.uid;
 405                 out->gid                = tbuf_old.msg_perm.gid;
 406                 out->mode               = tbuf_old.msg_perm.mode;
 407
 408                 if (tbuf_old.msg_qbytes == 0)
 409                         out->qbytes     = tbuf_old.msg_lqbytes;
 410                 else
 411                         out->qbytes     = tbuf_old.msg_qbytes;
 412
 413                 return 0;
 414         }
 415         default:
 416                 return -EINVAL;
 417         }
 418 }
 419
 420 asmlinkage long sys_msgctl(int msqid, int cmd, struct msqid_ds __user *buf)
 421 {
 422         struct kern_ipc_perm *ipcp;
 423         struct msq_setbuf uninitialized_var(setbuf);
 424         struct msg_queue *msq;
 425         int err, version;
 426         struct ipc_namespace *ns;
 427
 428         if (msqid < 0 || cmd < 0)
 429                 return -EINVAL;
 430
 431         version = ipc_parse_version(&cmd);
 432         ns = current->nsproxy->ipc_ns;
 433
 434         switch (cmd) {
 435         case IPC_INFO:
 436         case MSG_INFO:
 437         {
 438                 struct msginfo msginfo;
 439                 int max_id;
 440
 441                 if (!buf)
 442                         return -EFAULT;
 443                 /*
 444                  * We must not return kernel stack data.
 445                  * due to padding, it's not enough
 446                  * to set all member fields.
 447                  */
 448                 err = security_msg_queue_msgctl(NULL, cmd);
 449                 if (err)
 450                         return err;
 451
 452                 memset(&msginfo, 0, sizeof(msginfo));
 453                 msginfo.msgmni = ns->msg_ctlmni;
 454                 msginfo.msgmax = ns->msg_ctlmax;
 455                 msginfo.msgmnb = ns->msg_ctlmnb;
 456                 msginfo.msgssz = MSGSSZ;
 457                 msginfo.msgseg = MSGSEG;
 458                 down_read(&msg_ids(ns).rw_mutex);
 459                 if (cmd == MSG_INFO) {
 460                         msginfo.msgpool = msg_ids(ns).in_use;
 461                         msginfo.msgmap = atomic_read(&ns->msg_hdrs);
 462                         msginfo.msgtql = atomic_read(&ns->msg_bytes);
 463                 } else {
 464                         msginfo.msgmap = MSGMAP;
 465                         msginfo.msgpool = MSGPOOL;
 466                         msginfo.msgtql = MSGTQL;
 467                 }
 468                 max_id = ipc_get_maxid(&msg_ids(ns));
 469                 up_read(&msg_ids(ns).rw_mutex);
 470                 if (copy_to_user(buf, &msginfo, sizeof(struct msginfo)))
 471                         return -EFAULT;
 472                 return (max_id < 0) ? 0 : max_id;
 473         }
 474         case MSG_STAT:  /* msqid is an index rather than a msg queue id */
 475         case IPC_STAT:
 476         {
 477                 struct msqid64_ds tbuf;
 478                 int success_return;
 479
 480                 if (!buf)
 481                         return -EFAULT;
 482
 483                 if (cmd == MSG_STAT) {
 484                         msq = msg_lock(ns, msqid);
 485                         if (IS_ERR(msq))
 486                                 return PTR_ERR(msq);
 487                         success_return = msq->q_perm.id;
 488                 } else {
 489                         msq = msg_lock_check(ns, msqid);
 490                         if (IS_ERR(msq))
 491                                 return PTR_ERR(msq);
 492                         success_return = 0;
 493                 }
 494                 err = -EACCES;
 495                 if (ipcperms(&msq->q_perm, S_IRUGO))
 496                         goto out_unlock;
 497
 498                 err = security_msg_queue_msgctl(msq, cmd);
 499                 if (err)
 500                         goto out_unlock;
 501
 502                 memset(&tbuf, 0, sizeof(tbuf));
 503
 504                 kernel_to_ipc64_perm(&msq->q_perm, &tbuf.msg_perm);
 505                 tbuf.msg_stime  = msq->q_stime;
 506                 tbuf.msg_rtime  = msq->q_rtime;
 507                 tbuf.msg_ctime  = msq->q_ctime;
 508                 tbuf.msg_cbytes = msq->q_cbytes;
 509                 tbuf.msg_qnum   = msq->q_qnum;
 510                 tbuf.msg_qbytes = msq->q_qbytes;
 511                 tbuf.msg_lspid  = msq->q_lspid;
 512                 tbuf.msg_lrpid  = msq->q_lrpid;
 513                 msg_unlock(msq);
 514                 if (copy_msqid_to_user(buf, &tbuf, version))
 515                         return -EFAULT;
 516                 return success_return;
 517         }
 518         case IPC_SET:
 519                 if (!buf)
 520                         return -EFAULT;
 521                 if (copy_msqid_from_user(&setbuf, buf, version))
 522                         return -EFAULT;
 523                 break;
 524         case IPC_RMID:
 525                 break;
 526         default:
 527                 return  -EINVAL;
 528         }
 529
 530         down_write(&msg_ids(ns).rw_mutex);
 531         msq = msg_lock_check_down(ns, msqid);
 532         if (IS_ERR(msq)) {
 533                 err = PTR_ERR(msq);
 534                 goto out_up;
 535         }
 536
 537         ipcp = &msq->q_perm;
 538
 539         err = audit_ipc_obj(ipcp);
 540         if (err)
 541                 goto out_unlock_up;
 542         if (cmd == IPC_SET) {
 543                 err = audit_ipc_set_perm(setbuf.qbytes, setbuf.uid, setbuf.gid,
 544                                          setbuf.mode);
 545                 if (err)
 546                         goto out_unlock_up;
 547         }
 548
 549         err = -EPERM;
 550         if (current->euid != ipcp->cuid &&
 551             current->euid != ipcp->uid && !capable(CAP_SYS_ADMIN))
 552                 /* We _could_ check for CAP_CHOWN above, but we don't */
 553                 goto out_unlock_up;
 554
 555         err = security_msg_queue_msgctl(msq, cmd);
 556         if (err)
 557                 goto out_unlock_up;
 558
 559         switch (cmd) {
 560         case IPC_SET:
 561         {
 562                 err = -EPERM;
 563                 if (setbuf.qbytes > ns->msg_ctlmnb && !capable(CAP_SYS_RESOURCE))
 564                         goto out_unlock_up;
 565
 566                 msq->q_qbytes = setbuf.qbytes;
 567
 568                 ipcp->uid = setbuf.uid;
 569                 ipcp->gid = setbuf.gid;
 570                 ipcp->mode = (ipcp->mode & ~S_IRWXUGO) |
 571                              (S_IRWXUGO & setbuf.mode);
 572                 msq->q_ctime = get_seconds();
 573                 /* sleeping receivers might be excluded by
 574                  * stricter permissions.
 575                  */
 576                 expunge_all(msq, -EAGAIN);
 577                 /* sleeping senders might be able to send
 578                  * due to a larger queue size.
 579                  */
 580                 ss_wakeup(&msq->q_senders, 0);
 581                 msg_unlock(msq);
 582                 break;
 583         }
 584         case IPC_RMID:
 585                 freeque(ns, msq);
 586                 break;
 587         }
 588         err = 0;
 589 out_up:
 590         up_write(&msg_ids(ns).rw_mutex);
 591         return err;
 592 out_unlock_up:
 593         msg_unlock(msq);
 594         goto out_up;
 595 out_unlock:
 596         msg_unlock(msq);
 597         return err;
 598 }
 599
 600 static int testmsg(struct msg_msg *msg, long type, int mode)
 601 {
 602         switch(mode)
 603         {
 604                 case SEARCH_ANY:
 605                         return 1;
 606                 case SEARCH_LESSEQUAL:
 607                         if (msg->m_type <=type)
 608                                 return 1;
 609                         break;
 610                 case SEARCH_EQUAL:
 611                         if (msg->m_type == type)
 612                                 return 1;
 613                         break;
 614                 case SEARCH_NOTEQUAL:
 615                         if (msg->m_type != type)
 616                                 return 1;
 617                         break;
 618         }
 619         return 0;
 620 }
 621
 622 static inline int pipelined_send(struct msg_queue *msq, struct msg_msg *msg)
 623 {
 624         struct list_head *tmp;
 625
 626         tmp = msq->q_receivers.next;
 627         while (tmp != &msq->q_receivers) {
 628                 struct msg_receiver *msr;
 629
 630                 msr = list_entry(tmp, struct msg_receiver, r_list);
 631                 tmp = tmp->next;
 632                 if (testmsg(msg, msr->r_msgtype, msr->r_mode) &&
 633                     !security_msg_queue_msgrcv(msq, msg, msr->r_tsk,
 634                                                msr->r_msgtype, msr->r_mode)) {
 635
 636                         list_del(&msr->r_list);
 637                         if (msr->r_maxsize < msg->m_ts) {
 638                                 msr->r_msg = NULL;
 639                                 wake_up_process(msr->r_tsk);
 640                                 smp_mb();
 641                                 msr->r_msg = ERR_PTR(-E2BIG);
 642                         } else {
 643                                 msr->r_msg = NULL;
 644                                 msq->q_lrpid = task_pid_vnr(msr->r_tsk);
 645                                 msq->q_rtime = get_seconds();
 646                                 wake_up_process(msr->r_tsk);
 647                                 smp_mb();
 648                                 msr->r_msg = msg;
 649
 650                                 return 1;
 651                         }
 652                 }
 653         }
 654         return 0;
 655 }
 656
 657 long do_msgsnd(int msqid, long mtype, void __user *mtext,
 658                 size_t msgsz, int msgflg)
 659 {
 660         struct msg_queue *msq;
 661         struct msg_msg *msg;
 662         int err;
 663         struct ipc_namespace *ns;
 664
 665         ns = current->nsproxy->ipc_ns;
 666
 667         if (msgsz > ns->msg_ctlmax || (long) msgsz < 0 || msqid < 0)
 668                 return -EINVAL;
 669         if (mtype < 1)
 670                 return -EINVAL;
 671
 672         msg = load_msg(mtext, msgsz);
 673         if (IS_ERR(msg))
 674                 return PTR_ERR(msg);
 675
 676         msg->m_type = mtype;
 677         msg->m_ts = msgsz;
 678
 679         msq = msg_lock_check(ns, msqid);
 680         if (IS_ERR(msq)) {
 681                 err = PTR_ERR(msq);
 682                 goto out_free;
 683         }
 684
 685         for (;;) {
 686                 struct msg_sender s;
 687
 688                 err = -EACCES;
 689                 if (ipcperms(&msq->q_perm, S_IWUGO))
 690                         goto out_unlock_free;
 691
 692                 err = security_msg_queue_msgsnd(msq, msg, msgflg);
 693                 if (err)
 694                         goto out_unlock_free;
 695
 696                 if (msgsz + msq->q_cbytes <= msq->q_qbytes &&
 697                                 1 + msq->q_qnum <= msq->q_qbytes) {
 698                         break;
 699                 }
 700
 701                 /* queue full, wait: */
 702                 if (msgflg & IPC_NOWAIT) {
 703                         err = -EAGAIN;
 704                         goto out_unlock_free;
 705                 }
 706                 ss_add(msq, &s);
 707                 ipc_rcu_getref(msq);
 708                 msg_unlock(msq);
 709                 schedule();
 710
 711                 ipc_lock_by_ptr(&msq->q_perm);
 712                 ipc_rcu_putref(msq);
 713                 if (msq->q_perm.deleted) {
 714                         err = -EIDRM;
 715                         goto out_unlock_free;
 716                 }
 717                 ss_del(&s);
 718
 719                 if (signal_pending(current)) {
 720                         err = -ERESTARTNOHAND;
 721                         goto out_unlock_free;
 722                 }
 723         }
 724
 725         msq->q_lspid = task_tgid_vnr(current);
 726         msq->q_stime = get_seconds();
 727
 728         if (!pipelined_send(msq, msg)) {
 729                 /* noone is waiting for this message, enqueue it */
 730                 list_add_tail(&msg->m_list, &msq->q_messages);
 731                 msq->q_cbytes += msgsz;
 732                 msq->q_qnum++;
 733                 atomic_add(msgsz, &ns->msg_bytes);
 734                 atomic_inc(&ns->msg_hdrs);
 735         }
 736
 737         err = 0;
 738         msg = NULL;
 739
 740 out_unlock_free:
 741         msg_unlock(msq);
 742 out_free:
 743         if (msg != NULL)
 744                 free_msg(msg);
 745         return err;
 746 }
 747
 748 asmlinkage long
 749 sys_msgsnd(int msqid, struct msgbuf __user *msgp, size_t msgsz, int msgflg)
 750 {
 751         long mtype;
 752
 753         if (get_user(mtype, &msgp->mtype))
 754                 return -EFAULT;
 755         return do_msgsnd(msqid, mtype, msgp->mtext, msgsz, msgflg);
 756 }
 757
 758 static inline int convert_mode(long *msgtyp, int msgflg)
 759 {
 760         /*
 761          *  find message of correct type.
 762          *  msgtyp = 0 => get first.
 763          *  msgtyp > 0 => get first message of matching type.
 764          *  msgtyp < 0 => get message with least type must be < abs(msgtype).
 765          */
 766         if (*msgtyp == 0)
 767                 return SEARCH_ANY;
 768         if (*msgtyp < 0) {
 769                 *msgtyp = -*msgtyp;
 770                 return SEARCH_LESSEQUAL;
 771         }
 772         if (msgflg & MSG_EXCEPT)
 773                 return SEARCH_NOTEQUAL;
 774         return SEARCH_EQUAL;
 775 }
 776
 777 long do_msgrcv(int msqid, long *pmtype, void __user *mtext,
 778                 size_t msgsz, long msgtyp, int msgflg)
 779 {
 780         struct msg_queue *msq;
 781         struct msg_msg *msg;
 782         int mode;
 783         struct ipc_namespace *ns;
 784
 785         if (msqid < 0 || (long) msgsz < 0)
 786                 return -EINVAL;
 787         mode = convert_mode(&msgtyp, msgflg);
 788         ns = current->nsproxy->ipc_ns;
 789
 790         msq = msg_lock_check(ns, msqid);
 791         if (IS_ERR(msq))
 792                 return PTR_ERR(msq);
 793
 794         for (;;) {
 795                 struct msg_receiver msr_d;
 796                 struct list_head *tmp;
 797
 798                 msg = ERR_PTR(-EACCES);
 799                 if (ipcperms(&msq->q_perm, S_IRUGO))
 800                         goto out_unlock;
 801
 802                 msg = ERR_PTR(-EAGAIN);
 803                 tmp = msq->q_messages.next;
 804                 while (tmp != &msq->q_messages) {
 805                         struct msg_msg *walk_msg;
 806
 807                         walk_msg = list_entry(tmp, struct msg_msg, m_list);
 808                         if (testmsg(walk_msg, msgtyp, mode) &&
 809                             !security_msg_queue_msgrcv(msq, walk_msg, current,
 810                                                        msgtyp, mode)) {
 811
 812                                 msg = walk_msg;
 813                                 if (mode == SEARCH_LESSEQUAL &&
 814                                                 walk_msg->m_type != 1) {
 815                                         msg = walk_msg;
 816                                         msgtyp = walk_msg->m_type - 1;
 817                                 } else {
 818                                         msg = walk_msg;
 819                                         break;
 820                                 }
 821                         }
 822                         tmp = tmp->next;
 823                 }
 824                 if (!IS_ERR(msg)) {
 825                         /*
 826                          * Found a suitable message.
 827                          * Unlink it from the queue.
 828                          */
 829                         if ((msgsz < msg->m_ts) && !(msgflg & MSG_NOERROR)) {
 830                                 msg = ERR_PTR(-E2BIG);
 831                                 goto out_unlock;
 832                         }
 833                         list_del(&msg->m_list);
 834                         msq->q_qnum--;
 835                         msq->q_rtime = get_seconds();
 836                         msq->q_lrpid = task_tgid_vnr(current);
 837                         msq->q_cbytes -= msg->m_ts;
 838                         atomic_sub(msg->m_ts, &ns->msg_bytes);
 839                         atomic_dec(&ns->msg_hdrs);
 840                         ss_wakeup(&msq->q_senders, 0);
 841                         msg_unlock(msq);
 842                         break;
 843                 }
 844                 /* No message waiting. Wait for a message */
 845                 if (msgflg & IPC_NOWAIT) {
 846                         msg = ERR_PTR(-ENOMSG);
 847                         goto out_unlock;
 848                 }
 849                 list_add_tail(&msr_d.r_list, &msq->q_receivers);
 850                 msr_d.r_tsk = current;
 851                 msr_d.r_msgtype = msgtyp;
 852                 msr_d.r_mode = mode;
 853                 if (msgflg & MSG_NOERROR)
 854                         msr_d.r_maxsize = INT_MAX;
 855                 else
 856                         msr_d.r_maxsize = msgsz;
 857                 msr_d.r_msg = ERR_PTR(-EAGAIN);
 858                 current->state = TASK_INTERRUPTIBLE;
 859                 msg_unlock(msq);
 860
 861                 schedule();
 862
 863                 /* Lockless receive, part 1:
 864                  * Disable preemption.  We don't hold a reference to the queue
 865                  * and getting a reference would defeat the idea of a lockless
 866                  * operation, thus the code relies on rcu to guarantee the
 867                  * existance of msq:
 868                  * Prior to destruction, expunge_all(-EIRDM) changes r_msg.
 869                  * Thus if r_msg is -EAGAIN, then the queue not yet destroyed.
 870                  * rcu_read_lock() prevents preemption between reading r_msg
 871                  * and the spin_lock() inside ipc_lock_by_ptr().
 872                  */
 873                 rcu_read_lock();
 874
 875                 /* Lockless receive, part 2:
 876                  * Wait until pipelined_send or expunge_all are outside of
 877                  * wake_up_process(). There is a race with exit(), see
 878                  * ipc/mqueue.c for the details.
 879                  */
 880                 msg = (struct msg_msg*)msr_d.r_msg;
 881                 while (msg == NULL) {
 882                         cpu_relax();
 883                         msg = (struct msg_msg *)msr_d.r_msg;
 884                 }
 885
 886                 /* Lockless receive, part 3:
 887                  * If there is a message or an error then accept it without
 888                  * locking.
 889                  */
 890                 if (msg != ERR_PTR(-EAGAIN)) {
 891                         rcu_read_unlock();
 892                         break;
 893                 }
 894
 895                 /* Lockless receive, part 3:
 896                  * Acquire the queue spinlock.
 897                  */
 898                 ipc_lock_by_ptr(&msq->q_perm);
 899                 rcu_read_unlock();
 900
 901                 /* Lockless receive, part 4:
 902                  * Repeat test after acquiring the spinlock.
 903                  */
 904                 msg = (struct msg_msg*)msr_d.r_msg;
 905                 if (msg != ERR_PTR(-EAGAIN))
 906                         goto out_unlock;
 907
 908                 list_del(&msr_d.r_list);
 909                 if (signal_pending(current)) {
 910                         msg = ERR_PTR(-ERESTARTNOHAND);
 911 out_unlock:
 912                         msg_unlock(msq);
 913                         break;
 914                 }
 915         }
 916         if (IS_ERR(msg))
 917                 return PTR_ERR(msg);
 918
 919         msgsz = (msgsz > msg->m_ts) ? msg->m_ts : msgsz;
 920         *pmtype = msg->m_type;
 921         if (store_msg(mtext, msg, msgsz))
 922                 msgsz = -EFAULT;
 923
 924         free_msg(msg);
 925
 926         return msgsz;
 927 }
 928
 929 asmlinkage long sys_msgrcv(int msqid, struct msgbuf __user *msgp, size_t msgsz,
 930                            long msgtyp, int msgflg)
 931 {
 932         long err, mtype;
 933
 934         err =  do_msgrcv(msqid, &mtype, msgp->mtext, msgsz, msgtyp, msgflg);
 935         if (err < 0)
 936                 goto out;
 937
 938         if (put_user(mtype, &msgp->mtype))
 939                 err = -EFAULT;
 940 out:
 941         return err;
 942 }
 943
 944 #ifdef CONFIG_PROC_FS
 945 static int sysvipc_msg_proc_show(struct seq_file *s, void *it)
 946 {
 947         struct msg_queue *msq = it;
 948
 949         return seq_printf(s,
 950                         "%10d %10d  %4o  %10lu %10lu %5u %5u %5u %5u %5u %5u %10lu %10lu %10lu\n",
 951                         msq->q_perm.key,
 952                         msq->q_perm.id,
 953                         msq->q_perm.mode,
 954                         msq->q_cbytes,
 955                         msq->q_qnum,
 956                         msq->q_lspid,
 957                         msq->q_lrpid,
 958                         msq->q_perm.uid,
 959                         msq->q_perm.gid,
 960                         msq->q_perm.cuid,
 961                         msq->q_perm.cgid,
 962                         msq->q_stime,
 963                         msq->q_rtime,
 964                         msq->q_ctime);
 965 }
 966 #endif