mm: update my address
[linux-3.10.git] / ipc / sem.c
index 0b45a4d..fea0ad3 100644 (file)
--- a/ipc/sem.c
+++ b/ipc/sem.c
@@ -58,7 +58,7 @@
  * SMP-threaded, sysctl's added
  * (c) 1999 Manfred Spraul <manfred@colorfullife.com>
  * Enforced range limit on SEM_UNDO
- * (c) 2001 Red Hat Inc <alan@redhat.com>
+ * (c) 2001 Red Hat Inc
  * Lockless wakeup
  * (c) 2003 Manfred Spraul <manfred@colorfullife.com>
  *
@@ -91,7 +91,6 @@
 
 #define sem_unlock(sma)                ipc_unlock(&(sma)->sem_perm)
 #define sem_checkid(sma, semid)        ipc_checkid(&sma->sem_perm, semid)
-#define sem_buildid(id, seq)   ipc_buildid(id, seq)
 
 static int newary(struct ipc_namespace *, struct ipc_params *);
 static void freeary(struct ipc_namespace *, struct kern_ipc_perm *);
@@ -142,21 +141,6 @@ void __init sem_init (void)
 }
 
 /*
- * This routine is called in the paths where the rw_mutex is held to protect
- * access to the idr tree.
- */
-static inline struct sem_array *sem_lock_check_down(struct ipc_namespace *ns,
-                                               int id)
-{
-       struct kern_ipc_perm *ipcp = ipc_lock_check_down(&sem_ids(ns), id);
-
-       if (IS_ERR(ipcp))
-               return (struct sem_array *)ipcp;
-
-       return container_of(ipcp, struct sem_array, sem_perm);
-}
-
-/*
  * sem_lock_(check_) routines are called in the paths where the rw_mutex
  * is not held.
  */
@@ -181,6 +165,25 @@ static inline struct sem_array *sem_lock_check(struct ipc_namespace *ns,
        return container_of(ipcp, struct sem_array, sem_perm);
 }
 
+static inline void sem_lock_and_putref(struct sem_array *sma)
+{
+       ipc_lock_by_ptr(&sma->sem_perm);
+       ipc_rcu_putref(sma);
+}
+
+static inline void sem_getref_and_unlock(struct sem_array *sma)
+{
+       ipc_rcu_getref(sma);
+       ipc_unlock(&(sma)->sem_perm);
+}
+
+static inline void sem_putref(struct sem_array *sma)
+{
+       ipc_lock_by_ptr(&sma->sem_perm);
+       ipc_rcu_putref(sma);
+       ipc_unlock(&(sma)->sem_perm);
+}
+
 static inline void sem_rmid(struct ipc_namespace *ns, struct sem_array *s)
 {
        ipc_rmid(&sem_ids(ns), &s->sem_perm);
@@ -268,11 +271,9 @@ static int newary(struct ipc_namespace *ns, struct ipc_params *params)
        }
        ns->used_sems += nsems;
 
-       sma->sem_perm.id = sem_buildid(id, sma->sem_perm.seq);
        sma->sem_base = (struct sem *) &sma[1];
-       /* sma->sem_pending = NULL; */
-       sma->sem_pending_last = &sma->sem_pending;
-       /* sma->undo = NULL; */
+       INIT_LIST_HEAD(&sma->sem_pending);
+       INIT_LIST_HEAD(&sma->list_id);
        sma->sem_nsems = nsems;
        sma->sem_ctime = get_seconds();
        sem_unlock(sma);
@@ -329,38 +330,6 @@ asmlinkage long sys_semget(key_t key, int nsems, int semflg)
        return ipcget(ns, &sem_ids(ns), &sem_ops, &sem_params);
 }
 
-/* Manage the doubly linked list sma->sem_pending as a FIFO:
- * insert new queue elements at the tail sma->sem_pending_last.
- */
-static inline void append_to_queue (struct sem_array * sma,
-                                   struct sem_queue * q)
-{
-       *(q->prev = sma->sem_pending_last) = q;
-       *(sma->sem_pending_last = &q->next) = NULL;
-}
-
-static inline void prepend_to_queue (struct sem_array * sma,
-                                    struct sem_queue * q)
-{
-       q->next = sma->sem_pending;
-       *(q->prev = &sma->sem_pending) = q;
-       if (q->next)
-               q->next->prev = &q->next;
-       else /* sma->sem_pending_last == &sma->sem_pending */
-               sma->sem_pending_last = &q->next;
-}
-
-static inline void remove_from_queue (struct sem_array * sma,
-                                     struct sem_queue * q)
-{
-       *(q->prev) = q->next;
-       if (q->next)
-               q->next->prev = q->prev;
-       else /* sma->sem_pending_last == &q->next */
-               sma->sem_pending_last = q->prev;
-       q->prev = NULL; /* mark as removed */
-}
-
 /*
  * Determine whether a sequence of semaphore operations would succeed
  * all at once. Return 0 if yes, 1 if need to sleep, else return error code.
@@ -436,16 +405,15 @@ static void update_queue (struct sem_array * sma)
        int error;
        struct sem_queue * q;
 
-       q = sma->sem_pending;
-       while(q) {
+       q = list_entry(sma->sem_pending.next, struct sem_queue, list);
+       while (&q->list != &sma->sem_pending) {
                error = try_atomic_semop(sma, q->sops, q->nsops,
                                         q->undo, q->pid);
 
                /* Does q->sleeper still need to sleep? */
                if (error <= 0) {
                        struct sem_queue *n;
-                       remove_from_queue(sma,q);
-                       q->status = IN_WAKEUP;
+
                        /*
                         * Continue scanning. The next operation
                         * that must be checked depends on the type of the
@@ -456,11 +424,26 @@ static void update_queue (struct sem_array * sma)
                         *   for semaphore values to become 0.
                         * - if the operation didn't modify the array,
                         *   then just continue.
+                        * The order of list_del() and reading ->next
+                        * is crucial: In the former case, the list_del()
+                        * must be done first [because we might be the
+                        * first entry in ->sem_pending], in the latter
+                        * case the list_del() must be done last
+                        * [because the list is invalid after the list_del()]
                         */
-                       if (q->alter)
-                               n = sma->sem_pending;
-                       else
-                               n = q->next;
+                       if (q->alter) {
+                               list_del(&q->list);
+                               n = list_entry(sma->sem_pending.next,
+                                               struct sem_queue, list);
+                       } else {
+                               n = list_entry(q->list.next, struct sem_queue,
+                                               list);
+                               list_del(&q->list);
+                       }
+
+                       /* wake up the waiting thread */
+                       q->status = IN_WAKEUP;
+
                        wake_up_process(q->sleeper);
                        /* hands-off: q will disappear immediately after
                         * writing q->status.
@@ -469,7 +452,7 @@ static void update_queue (struct sem_array * sma)
                        q->status = error;
                        q = n;
                } else {
-                       q = q->next;
+                       q = list_entry(q->list.next, struct sem_queue, list);
                }
        }
 }
@@ -489,7 +472,7 @@ static int count_semncnt (struct sem_array * sma, ushort semnum)
        struct sem_queue * q;
 
        semncnt = 0;
-       for (q = sma->sem_pending; q; q = q->next) {
+       list_for_each_entry(q, &sma->sem_pending, list) {
                struct sembuf * sops = q->sops;
                int nsops = q->nsops;
                int i;
@@ -501,13 +484,14 @@ static int count_semncnt (struct sem_array * sma, ushort semnum)
        }
        return semncnt;
 }
+
 static int count_semzcnt (struct sem_array * sma, ushort semnum)
 {
        int semzcnt;
        struct sem_queue * q;
 
        semzcnt = 0;
-       for (q = sma->sem_pending; q; q = q->next) {
+       list_for_each_entry(q, &sma->sem_pending, list) {
                struct sembuf * sops = q->sops;
                int nsops = q->nsops;
                int i;
@@ -520,35 +504,41 @@ static int count_semzcnt (struct sem_array * sma, ushort semnum)
        return semzcnt;
 }
 
+static void free_un(struct rcu_head *head)
+{
+       struct sem_undo *un = container_of(head, struct sem_undo, rcu);
+       kfree(un);
+}
+
 /* Free a semaphore set. freeary() is called with sem_ids.rw_mutex locked
  * as a writer and the spinlock for this semaphore set hold. sem_ids.rw_mutex
  * remains locked on exit.
  */
 static void freeary(struct ipc_namespace *ns, struct kern_ipc_perm *ipcp)
 {
-       struct sem_undo *un;
-       struct sem_queue *q;
+       struct sem_undo *un, *tu;
+       struct sem_queue *q, *tq;
        struct sem_array *sma = container_of(ipcp, struct sem_array, sem_perm);
 
-       /* Invalidate the existing undo structures for this semaphore set.
-        * (They will be freed without any further action in exit_sem()
-        * or during the next semop.)
-        */
-       for (un = sma->undo; un; un = un->id_next)
+       /* Free the existing undo structures for this semaphore set.  */
+       assert_spin_locked(&sma->sem_perm.lock);
+       list_for_each_entry_safe(un, tu, &sma->list_id, list_id) {
+               list_del(&un->list_id);
+               spin_lock(&un->ulp->lock);
                un->semid = -1;
+               list_del_rcu(&un->list_proc);
+               spin_unlock(&un->ulp->lock);
+               call_rcu(&un->rcu, free_un);
+       }
 
        /* Wake up all pending processes and let them fail with EIDRM. */
-       q = sma->sem_pending;
-       while(q) {
-               struct sem_queue *n;
-               /* lazy remove_from_queue: we are killing the whole queue */
-               q->prev = NULL;
-               n = q->next;
+       list_for_each_entry_safe(q, tq, &sma->sem_pending, list) {
+               list_del(&q->list);
+
                q->status = IN_WAKEUP;
                wake_up_process(q->sleeper); /* doesn't sleep */
                smp_wmb();
                q->status = -EIDRM;     /* hands-off q */
-               q = n;
        }
 
        /* Remove the semaphore set from the IDR */
@@ -700,19 +690,15 @@ static int semctl_main(struct ipc_namespace *ns, int semid, int semnum,
                int i;
 
                if(nsems > SEMMSL_FAST) {
-                       ipc_rcu_getref(sma);
-                       sem_unlock(sma);                        
+                       sem_getref_and_unlock(sma);
 
                        sem_io = ipc_alloc(sizeof(ushort)*nsems);
                        if(sem_io == NULL) {
-                               ipc_lock_by_ptr(&sma->sem_perm);
-                               ipc_rcu_putref(sma);
-                               sem_unlock(sma);
+                               sem_putref(sma);
                                return -ENOMEM;
                        }
 
-                       ipc_lock_by_ptr(&sma->sem_perm);
-                       ipc_rcu_putref(sma);
+                       sem_lock_and_putref(sma);
                        if (sma->sem_perm.deleted) {
                                sem_unlock(sma);
                                err = -EIDRM;
@@ -733,38 +719,30 @@ static int semctl_main(struct ipc_namespace *ns, int semid, int semnum,
                int i;
                struct sem_undo *un;
 
-               ipc_rcu_getref(sma);
-               sem_unlock(sma);
+               sem_getref_and_unlock(sma);
 
                if(nsems > SEMMSL_FAST) {
                        sem_io = ipc_alloc(sizeof(ushort)*nsems);
                        if(sem_io == NULL) {
-                               ipc_lock_by_ptr(&sma->sem_perm);
-                               ipc_rcu_putref(sma);
-                               sem_unlock(sma);
+                               sem_putref(sma);
                                return -ENOMEM;
                        }
                }
 
                if (copy_from_user (sem_io, arg.array, nsems*sizeof(ushort))) {
-                       ipc_lock_by_ptr(&sma->sem_perm);
-                       ipc_rcu_putref(sma);
-                       sem_unlock(sma);
+                       sem_putref(sma);
                        err = -EFAULT;
                        goto out_free;
                }
 
                for (i = 0; i < nsems; i++) {
                        if (sem_io[i] > SEMVMX) {
-                               ipc_lock_by_ptr(&sma->sem_perm);
-                               ipc_rcu_putref(sma);
-                               sem_unlock(sma);
+                               sem_putref(sma);
                                err = -ERANGE;
                                goto out_free;
                        }
                }
-               ipc_lock_by_ptr(&sma->sem_perm);
-               ipc_rcu_putref(sma);
+               sem_lock_and_putref(sma);
                if (sma->sem_perm.deleted) {
                        sem_unlock(sma);
                        err = -EIDRM;
@@ -773,9 +751,12 @@ static int semctl_main(struct ipc_namespace *ns, int semid, int semnum,
 
                for (i = 0; i < nsems; i++)
                        sma->sem_base[i].semval = sem_io[i];
-               for (un = sma->undo; un; un = un->id_next)
+
+               assert_spin_locked(&sma->sem_perm.lock);
+               list_for_each_entry(un, &sma->list_id, list_id) {
                        for (i = 0; i < nsems; i++)
                                un->semadj[i] = 0;
+               }
                sma->sem_ctime = get_seconds();
                /* maybe some queued-up processes were waiting for this */
                update_queue(sma);
@@ -807,12 +788,15 @@ static int semctl_main(struct ipc_namespace *ns, int semid, int semnum,
        {
                int val = arg.val;
                struct sem_undo *un;
+
                err = -ERANGE;
                if (val > SEMVMX || val < 0)
                        goto out_unlock;
 
-               for (un = sma->undo; un; un = un->id_next)
+               assert_spin_locked(&sma->sem_perm.lock);
+               list_for_each_entry(un, &sma->list_id, list_id)
                        un->semadj[semnum] = 0;
+
                curr->semval = val;
                curr->sempid = task_tgid_vnr(current);
                sma->sem_ctime = get_seconds();
@@ -830,28 +814,14 @@ out_free:
        return err;
 }
 
-struct sem_setbuf {
-       uid_t   uid;
-       gid_t   gid;
-       mode_t  mode;
-};
-
-static inline unsigned long copy_semid_from_user(struct sem_setbuf *out, void __user *buf, int version)
+static inline unsigned long
+copy_semid_from_user(struct semid64_ds *out, void __user *buf, int version)
 {
        switch(version) {
        case IPC_64:
-           {
-               struct semid64_ds tbuf;
-
-               if(copy_from_user(&tbuf, buf, sizeof(tbuf)))
+               if (copy_from_user(out, buf, sizeof(*out)))
                        return -EFAULT;
-
-               out->uid        = tbuf.sem_perm.uid;
-               out->gid        = tbuf.sem_perm.gid;
-               out->mode       = tbuf.sem_perm.mode;
-
                return 0;
-           }
        case IPC_OLD:
            {
                struct semid_ds tbuf_old;
@@ -859,9 +829,9 @@ static inline unsigned long copy_semid_from_user(struct sem_setbuf *out, void __
                if(copy_from_user(&tbuf_old, buf, sizeof(tbuf_old)))
                        return -EFAULT;
 
-               out->uid        = tbuf_old.sem_perm.uid;
-               out->gid        = tbuf_old.sem_perm.gid;
-               out->mode       = tbuf_old.sem_perm.mode;
+               out->sem_perm.uid       = tbuf_old.sem_perm.uid;
+               out->sem_perm.gid       = tbuf_old.sem_perm.gid;
+               out->sem_perm.mode      = tbuf_old.sem_perm.mode;
 
                return 0;
            }
@@ -870,38 +840,29 @@ static inline unsigned long copy_semid_from_user(struct sem_setbuf *out, void __
        }
 }
 
-static int semctl_down(struct ipc_namespace *ns, int semid, int semnum,
-               int cmd, int version, union semun arg)
+/*
+ * This function handles some semctl commands which require the rw_mutex
+ * to be held in write mode.
+ * NOTE: no locks must be held, the rw_mutex is taken inside this function.
+ */
+static int semctl_down(struct ipc_namespace *ns, int semid,
+                      int cmd, int version, union semun arg)
 {
        struct sem_array *sma;
        int err;
-       struct sem_setbuf uninitialized_var(setbuf);
+       struct semid64_ds semid64;
        struct kern_ipc_perm *ipcp;
 
        if(cmd == IPC_SET) {
-               if(copy_semid_from_user (&setbuf, arg.buf, version))
+               if (copy_semid_from_user(&semid64, arg.buf, version))
                        return -EFAULT;
        }
-       sma = sem_lock_check_down(ns, semid);
-       if (IS_ERR(sma))
-               return PTR_ERR(sma);
-
-       ipcp = &sma->sem_perm;
 
-       err = audit_ipc_obj(ipcp);
-       if (err)
-               goto out_unlock;
+       ipcp = ipcctl_pre_down(&sem_ids(ns), semid, cmd, &semid64.sem_perm, 0);
+       if (IS_ERR(ipcp))
+               return PTR_ERR(ipcp);
 
-       if (cmd == IPC_SET) {
-               err = audit_ipc_set_perm(0, setbuf.uid, setbuf.gid, setbuf.mode);
-               if (err)
-                       goto out_unlock;
-       }
-       if (current->euid != ipcp->cuid && 
-           current->euid != ipcp->uid && !capable(CAP_SYS_ADMIN)) {
-               err=-EPERM;
-               goto out_unlock;
-       }
+       sma = container_of(ipcp, struct sem_array, sem_perm);
 
        err = security_sem_semctl(sma, cmd);
        if (err)
@@ -910,26 +871,19 @@ static int semctl_down(struct ipc_namespace *ns, int semid, int semnum,
        switch(cmd){
        case IPC_RMID:
                freeary(ns, ipcp);
-               err = 0;
-               break;
+               goto out_up;
        case IPC_SET:
-               ipcp->uid = setbuf.uid;
-               ipcp->gid = setbuf.gid;
-               ipcp->mode = (ipcp->mode & ~S_IRWXUGO)
-                               | (setbuf.mode & S_IRWXUGO);
+               ipc_update_perm(&semid64.sem_perm, ipcp);
                sma->sem_ctime = get_seconds();
-               sem_unlock(sma);
-               err = 0;
                break;
        default:
-               sem_unlock(sma);
                err = -EINVAL;
-               break;
        }
-       return err;
 
 out_unlock:
        sem_unlock(sma);
+out_up:
+       up_write(&sem_ids(ns).rw_mutex);
        return err;
 }
 
@@ -963,9 +917,7 @@ asmlinkage long sys_semctl (int semid, int semnum, int cmd, union semun arg)
                return err;
        case IPC_RMID:
        case IPC_SET:
-               down_write(&sem_ids(ns).rw_mutex);
-               err = semctl_down(ns,semid,semnum,cmd,version,arg);
-               up_write(&sem_ids(ns).rw_mutex);
+               err = semctl_down(ns, semid, cmd, version, arg);
                return err;
        default:
                return -EINVAL;
@@ -994,6 +946,8 @@ static inline int get_undo_list(struct sem_undo_list **undo_listp)
                        return -ENOMEM;
                spin_lock_init(&undo_list->lock);
                atomic_set(&undo_list->refcnt, 1);
+               INIT_LIST_HEAD(&undo_list->list_proc);
+
                current->sysvsem.undo_list = undo_list;
        }
        *undo_listp = undo_list;
@@ -1002,25 +956,27 @@ static inline int get_undo_list(struct sem_undo_list **undo_listp)
 
 static struct sem_undo *lookup_undo(struct sem_undo_list *ulp, int semid)
 {
-       struct sem_undo **last, *un;
+       struct sem_undo *walk;
 
-       last = &ulp->proc_list;
-       un = *last;
-       while(un != NULL) {
-               if(un->semid==semid)
-                       break;
-               if(un->semid==-1) {
-                       *last=un->proc_next;
-                       kfree(un);
-               } else {
-                       last=&un->proc_next;
-               }
-               un=*last;
+       list_for_each_entry_rcu(walk, &ulp->list_proc, list_proc) {
+               if (walk->semid == semid)
+                       return walk;
        }
-       return un;
+       return NULL;
 }
 
-static struct sem_undo *find_undo(struct ipc_namespace *ns, int semid)
+/**
+ * find_alloc_undo - Lookup (and if not present create) undo array
+ * @ns: namespace
+ * @semid: semaphore array id
+ *
+ * The function looks up (and if not present creates) the undo structure.
+ * The size of the undo structure depends on the size of the semaphore
+ * array, thus the alloc path is not that straightforward.
+ * Lifetime-rules: sem_undo is rcu-protected, on success, the function
+ * performs a rcu_read_lock().
+ */
+static struct sem_undo *find_alloc_undo(struct ipc_namespace *ns, int semid)
 {
        struct sem_array *sma;
        struct sem_undo_list *ulp;
@@ -1032,57 +988,62 @@ static struct sem_undo *find_undo(struct ipc_namespace *ns, int semid)
        if (error)
                return ERR_PTR(error);
 
+       rcu_read_lock();
        spin_lock(&ulp->lock);
        un = lookup_undo(ulp, semid);
        spin_unlock(&ulp->lock);
        if (likely(un!=NULL))
                goto out;
+       rcu_read_unlock();
 
        /* no undo structure around - allocate one. */
+       /* step 1: figure out the size of the semaphore array */
        sma = sem_lock_check(ns, semid);
        if (IS_ERR(sma))
                return ERR_PTR(PTR_ERR(sma));
 
        nsems = sma->sem_nsems;
-       ipc_rcu_getref(sma);
-       sem_unlock(sma);
+       sem_getref_and_unlock(sma);
 
+       /* step 2: allocate new undo structure */
        new = kzalloc(sizeof(struct sem_undo) + sizeof(short)*nsems, GFP_KERNEL);
        if (!new) {
-               ipc_lock_by_ptr(&sma->sem_perm);
-               ipc_rcu_putref(sma);
-               sem_unlock(sma);
+               sem_putref(sma);
                return ERR_PTR(-ENOMEM);
        }
-       new->semadj = (short *) &new[1];
-       new->semid = semid;
 
-       spin_lock(&ulp->lock);
-       un = lookup_undo(ulp, semid);
-       if (un) {
-               spin_unlock(&ulp->lock);
-               kfree(new);
-               ipc_lock_by_ptr(&sma->sem_perm);
-               ipc_rcu_putref(sma);
-               sem_unlock(sma);
-               goto out;
-       }
-       ipc_lock_by_ptr(&sma->sem_perm);
-       ipc_rcu_putref(sma);
+       /* step 3: Acquire the lock on semaphore array */
+       sem_lock_and_putref(sma);
        if (sma->sem_perm.deleted) {
                sem_unlock(sma);
-               spin_unlock(&ulp->lock);
                kfree(new);
                un = ERR_PTR(-EIDRM);
                goto out;
        }
-       new->proc_next = ulp->proc_list;
-       ulp->proc_list = new;
-       new->id_next = sma->undo;
-       sma->undo = new;
-       sem_unlock(sma);
+       spin_lock(&ulp->lock);
+
+       /*
+        * step 4: check for races: did someone else allocate the undo struct?
+        */
+       un = lookup_undo(ulp, semid);
+       if (un) {
+               kfree(new);
+               goto success;
+       }
+       /* step 5: initialize & link new undo structure */
+       new->semadj = (short *) &new[1];
+       new->ulp = ulp;
+       new->semid = semid;
+       assert_spin_locked(&ulp->lock);
+       list_add_rcu(&new->list_proc, &ulp->list_proc);
+       assert_spin_locked(&sma->sem_perm.lock);
+       list_add(&new->list_id, &sma->list_id);
        un = new;
+
+success:
        spin_unlock(&ulp->lock);
+       rcu_read_lock();
+       sem_unlock(sma);
 out:
        return un;
 }
@@ -1138,9 +1099,8 @@ asmlinkage long sys_semtimedop(int semid, struct sembuf __user *tsops,
                        alter = 1;
        }
 
-retry_undos:
        if (undos) {
-               un = find_undo(ns, semid);
+               un = find_alloc_undo(ns, semid);
                if (IS_ERR(un)) {
                        error = PTR_ERR(un);
                        goto out_free;
@@ -1150,19 +1110,37 @@ retry_undos:
 
        sma = sem_lock_check(ns, semid);
        if (IS_ERR(sma)) {
+               if (un)
+                       rcu_read_unlock();
                error = PTR_ERR(sma);
                goto out_free;
        }
 
        /*
-        * semid identifiers are not unique - find_undo may have
+        * semid identifiers are not unique - find_alloc_undo may have
         * allocated an undo structure, it was invalidated by an RMID
-        * and now a new array with received the same id. Check and retry.
+        * and now a new array with received the same id. Check and fail.
+        * This case can be detected checking un->semid. The existance of
+        * "un" itself is guaranteed by rcu.
         */
-       if (un && un->semid == -1) {
-               sem_unlock(sma);
-               goto retry_undos;
+       error = -EIDRM;
+       if (un) {
+               if (un->semid == -1) {
+                       rcu_read_unlock();
+                       goto out_unlock_free;
+               } else {
+                       /*
+                        * rcu lock can be released, "un" cannot disappear:
+                        * - sem_lock is acquired, thus IPC_RMID is
+                        *   impossible.
+                        * - exit_sem is impossible, it always operates on
+                        *   current (or a dead task).
+                        */
+
+                       rcu_read_unlock();
+               }
        }
+
        error = -EFBIG;
        if (max >= sma->sem_nsems)
                goto out_unlock_free;
@@ -1186,17 +1164,15 @@ retry_undos:
         * task into the pending queue and go to sleep.
         */
                
-       queue.sma = sma;
        queue.sops = sops;
        queue.nsops = nsops;
        queue.undo = un;
        queue.pid = task_tgid_vnr(current);
-       queue.id = semid;
        queue.alter = alter;
        if (alter)
-               append_to_queue(sma ,&queue);
+               list_add_tail(&queue.list, &sma->sem_pending);
        else
-               prepend_to_queue(sma ,&queue);
+               list_add(&queue.list, &sma->sem_pending);
 
        queue.status = -EINTR;
        queue.sleeper = current;
@@ -1222,7 +1198,6 @@ retry_undos:
 
        sma = sem_lock(ns, semid);
        if (IS_ERR(sma)) {
-               BUG_ON(queue.prev != NULL);
                error = -EIDRM;
                goto out_free;
        }
@@ -1240,7 +1215,7 @@ retry_undos:
         */
        if (timeout && jiffies_left == 0)
                error = -EAGAIN;
-       remove_from_queue(sma,&queue);
+       list_del(&queue.list);
        goto out_unlock_free;
 
 out_unlock_free:
@@ -1291,55 +1266,62 @@ int copy_semundo(unsigned long clone_flags, struct task_struct *tsk)
  */
 void exit_sem(struct task_struct *tsk)
 {
-       struct sem_undo_list *undo_list;
-       struct sem_undo *u, **up;
-       struct ipc_namespace *ns;
+       struct sem_undo_list *ulp;
 
-       undo_list = tsk->sysvsem.undo_list;
-       if (!undo_list)
+       ulp = tsk->sysvsem.undo_list;
+       if (!ulp)
                return;
+       tsk->sysvsem.undo_list = NULL;
 
-       if (!atomic_dec_and_test(&undo_list->refcnt))
+       if (!atomic_dec_and_test(&ulp->refcnt))
                return;
 
-       ns = tsk->nsproxy->ipc_ns;
-       /* There's no need to hold the semundo list lock, as current
-         * is the last task exiting for this undo list.
-        */
-       for (up = &undo_list->proc_list; (u = *up); *up = u->proc_next, kfree(u)) {
+       for (;;) {
                struct sem_array *sma;
-               int nsems, i;
-               struct sem_undo *un, **unp;
+               struct sem_undo *un;
                int semid;
-              
-               semid = u->semid;
+               int i;
 
-               if(semid == -1)
-                       continue;
-               sma = sem_lock(ns, semid);
+               rcu_read_lock();
+               un = list_entry(rcu_dereference(ulp->list_proc.next),
+                                       struct sem_undo, list_proc);
+               if (&un->list_proc == &ulp->list_proc)
+                       semid = -1;
+                else
+                       semid = un->semid;
+               rcu_read_unlock();
+
+               if (semid == -1)
+                       break;
+
+               sma = sem_lock_check(tsk->nsproxy->ipc_ns, un->semid);
+
+               /* exit_sem raced with IPC_RMID, nothing to do */
                if (IS_ERR(sma))
                        continue;
 
-               if (u->semid == -1)
-                       goto next_entry;
+               un = lookup_undo(ulp, semid);
+               if (un == NULL) {
+                       /* exit_sem raced with IPC_RMID+semget() that created
+                        * exactly the same semid. Nothing to do.
+                        */
+                       sem_unlock(sma);
+                       continue;
+               }
 
-               BUG_ON(sem_checkid(sma, u->semid));
+               /* remove un from the linked lists */
+               assert_spin_locked(&sma->sem_perm.lock);
+               list_del(&un->list_id);
 
-               /* remove u from the sma->undo list */
-               for (unp = &sma->undo; (un = *unp); unp = &un->id_next) {
-                       if (u == un)
-                               goto found;
-               }
-               printk ("exit_sem undo list error id=%d\n", u->semid);
-               goto next_entry;
-found:
-               *unp = un->id_next;
-               /* perform adjustments registered in u */
-               nsems = sma->sem_nsems;
-               for (i = 0; i < nsems; i++) {
+               spin_lock(&ulp->lock);
+               list_del_rcu(&un->list_proc);
+               spin_unlock(&ulp->lock);
+
+               /* perform adjustments registered in un */
+               for (i = 0; i < sma->sem_nsems; i++) {
                        struct sem * semaphore = &sma->sem_base[i];
-                       if (u->semadj[i]) {
-                               semaphore->semval += u->semadj[i];
+                       if (un->semadj[i]) {
+                               semaphore->semval += un->semadj[i];
                                /*
                                 * Range checks of the new semaphore value,
                                 * not defined by sus:
@@ -1363,10 +1345,11 @@ found:
                sma->sem_otime = get_seconds();
                /* maybe some queued-up processes were waiting for this */
                update_queue(sma);
-next_entry:
                sem_unlock(sma);
+
+               call_rcu(&un->rcu, free_un);
        }
-       kfree(undo_list);
+       kfree(ulp);
 }
 
 #ifdef CONFIG_PROC_FS